6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subtypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "utilities/bitMap.inline.hpp"
50 #include "utilities/growableArray.hpp"
51 #include "utilities/powerOfTwo.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn()),
59 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
60 {
61 _exceptions = jvms->map()->next_exception();
62 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
63 set_jvms(jvms);
64 }
65
66 // Private constructor for parser.
67 GraphKit::GraphKit()
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn(*C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 _exceptions = nullptr;
74 set_map(nullptr);
75 DEBUG_ONLY(_sp = -99);
76 DEBUG_ONLY(set_bci(-99));
77 }
78
79
80
81 //---------------------------clean_stack---------------------------------------
82 // Clear away rubbish from the stack area of the JVM state.
83 // This destroys any arguments that may be waiting on the stack.
84 void GraphKit::clean_stack(int from_sp) {
85 SafePointNode* map = this->map();
86 JVMState* jvms = this->jvms();
87 int stk_size = jvms->stk_size();
88 int stkoff = jvms->stkoff();
89 Node* top = this->top();
90 for (int i = from_sp; i < stk_size; i++) {
91 if (map->in(stkoff + i) != top) {
92 map->set_req(stkoff + i, top);
93 }
94 }
95 }
96
97
98 //--------------------------------sync_jvms-----------------------------------
99 // Make sure our current jvms agrees with our parse state.
328 }
329 static inline void add_one_req(Node* dstphi, Node* src) {
330 assert(is_hidden_merge(dstphi), "must be a special merge node");
331 assert(!is_hidden_merge(src), "must not be a special merge node");
332 dstphi->add_req(src);
333 }
334
335 //-----------------------combine_exception_states------------------------------
336 // This helper function combines exception states by building phis on a
337 // specially marked state-merging region. These regions and phis are
338 // untransformed, and can build up gradually. The region is marked by
339 // having a control input of its exception map, rather than null. Such
340 // regions do not appear except in this function, and in use_exception_state.
341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
342 if (failing_internal()) {
343 return; // dying anyway...
344 }
345 JVMState* ex_jvms = ex_map->_jvms;
346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
351 assert(ex_map->req() == phi_map->req(), "matching maps");
352 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
353 Node* hidden_merge_mark = root();
354 Node* region = phi_map->control();
355 MergeMemNode* phi_mem = phi_map->merged_memory();
356 MergeMemNode* ex_mem = ex_map->merged_memory();
357 if (region->in(0) != hidden_merge_mark) {
358 // The control input is not (yet) a specially-marked region in phi_map.
359 // Make it so, and build some phis.
360 region = new RegionNode(2);
361 _gvn.set_type(region, Type::CONTROL);
362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
363 region->init_req(1, phi_map->control());
364 phi_map->set_control(region);
365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
366 record_for_igvn(io_phi);
367 _gvn.set_type(io_phi, Type::ABIO);
368 phi_map->set_i_o(io_phi);
898 if (PrintMiscellaneous && (Verbose || WizardMode)) {
899 tty->print_cr("Zombie local %d: ", local);
900 jvms->dump();
901 }
902 return false;
903 }
904 }
905 }
906 return true;
907 }
908
909 #endif //ASSERT
910
911 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
912 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
913 ciMethod* cur_method = jvms->method();
914 int cur_bci = jvms->bci();
915 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
916 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
917 return Interpreter::bytecode_should_reexecute(code) ||
918 (is_anewarray && code == Bytecodes::_multianewarray);
919 // Reexecute _multianewarray bytecode which was replaced with
920 // sequence of [a]newarray. See Parse::do_multianewarray().
921 //
922 // Note: interpreter should not have it set since this optimization
923 // is limited by dimensions and guarded by flag so in some cases
924 // multianewarray() runtime calls will be generated and
925 // the bytecode should not be reexecutes (stack will not be reset).
926 } else {
927 return false;
928 }
929 }
930
931 // Helper function for adding JVMState and debug information to node
932 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
933 // Add the safepoint edges to the call (or other safepoint).
934
935 // Make sure dead locals are set to top. This
936 // should help register allocation time and cut down on the size
937 // of the deoptimization information.
938 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
966
967 if (env()->should_retain_local_variables()) {
968 // At any safepoint, this method can get breakpointed, which would
969 // then require an immediate deoptimization.
970 can_prune_locals = false; // do not prune locals
971 stack_slots_not_pruned = 0;
972 }
973
974 // do not scribble on the input jvms
975 JVMState* out_jvms = youngest_jvms->clone_deep(C);
976 call->set_jvms(out_jvms); // Start jvms list for call node
977
978 // For a known set of bytecodes, the interpreter should reexecute them if
979 // deoptimization happens. We set the reexecute state for them here
980 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
981 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
982 #ifdef ASSERT
983 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
984 assert(method() == youngest_jvms->method(), "sanity");
985 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
986 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
987 #endif // ASSERT
988 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
989 }
990
991 // Presize the call:
992 DEBUG_ONLY(uint non_debug_edges = call->req());
993 call->add_req_batch(top(), youngest_jvms->debug_depth());
994 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
995
996 // Set up edges so that the call looks like this:
997 // Call [state:] ctl io mem fptr retadr
998 // [parms:] parm0 ... parmN
999 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1000 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
1001 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1002 // Note that caller debug info precedes callee debug info.
1003
1004 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1005 uint debug_ptr = call->req();
1006
1007 // Loop over the map input edges associated with jvms, add them
1008 // to the call node, & reset all offsets to match call node array.
1009 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1010 uint debug_end = debug_ptr;
1011 uint debug_start = debug_ptr - in_jvms->debug_size();
1012 debug_ptr = debug_start; // back up the ptr
1013
1014 uint p = debug_start; // walks forward in [debug_start, debug_end)
1015 uint j, k, l;
1016 SafePointNode* in_map = in_jvms->map();
1017 out_jvms->set_map(call);
1018
1019 if (can_prune_locals) {
1020 assert(in_jvms->method() == out_jvms->method(), "sanity");
1021 // If the current throw can reach an exception handler in this JVMS,
1022 // then we must keep everything live that can reach that handler.
1023 // As a quick and dirty approximation, we look for any handlers at all.
1024 if (in_jvms->method()->has_exception_handlers()) {
1025 can_prune_locals = false;
1026 }
1027 }
1028
1029 // Add the Locals
1030 k = in_jvms->locoff();
1031 l = in_jvms->loc_size();
1032 out_jvms->set_locoff(p);
1033 if (!can_prune_locals) {
1034 for (j = 0; j < l; j++)
1035 call->set_req(p++, in_map->in(k+j));
1036 } else {
1037 p += l; // already set to top above by add_req_batch
1038 }
1039
1040 // Add the Expression Stack
1041 k = in_jvms->stkoff();
1042 l = in_jvms->sp();
1043 out_jvms->set_stkoff(p);
1044 if (!can_prune_locals) {
1045 for (j = 0; j < l; j++)
1046 call->set_req(p++, in_map->in(k+j));
1047 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1048 // Divide stack into {S0,...,S1}, where S0 is set to top.
1049 uint s1 = stack_slots_not_pruned;
1050 stack_slots_not_pruned = 0; // for next iteration
1051 if (s1 > l) s1 = l;
1052 uint s0 = l - s1;
1053 p += s0; // skip the tops preinstalled by add_req_batch
1054 for (j = s0; j < l; j++)
1055 call->set_req(p++, in_map->in(k+j));
1056 } else {
1057 p += l; // already set to top above by add_req_batch
1058 }
1059
1060 // Add the Monitors
1061 k = in_jvms->monoff();
1062 l = in_jvms->mon_size();
1063 out_jvms->set_monoff(p);
1064 for (j = 0; j < l; j++)
1065 call->set_req(p++, in_map->in(k+j));
1066
1067 // Copy any scalar object fields.
1068 k = in_jvms->scloff();
1069 l = in_jvms->scl_size();
1070 out_jvms->set_scloff(p);
1071 for (j = 0; j < l; j++)
1072 call->set_req(p++, in_map->in(k+j));
1073
1074 // Finish the new jvms.
1075 out_jvms->set_endoff(p);
1076
1077 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1078 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1079 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1080 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1081 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1082 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1083
1084 // Update the two tail pointers in parallel.
1085 out_jvms = out_jvms->caller();
1086 in_jvms = in_jvms->caller();
1087 }
1088
1089 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1090
1091 // Test the correctness of JVMState::debug_xxx accessors:
1092 assert(call->jvms()->debug_start() == non_debug_edges, "");
1093 assert(call->jvms()->debug_end() == call->req(), "");
1094 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1095 }
1096
1097 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1098 Bytecodes::Code code = java_bc();
1099 if (code == Bytecodes::_wide) {
1100 code = method()->java_code_at_bci(bci() + 1);
1101 }
1102
1103 if (code != Bytecodes::_illegal) {
1104 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1254 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1255 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1256 return _gvn.transform( new AndLNode(conv, mask) );
1257 }
1258
1259 Node* GraphKit::ConvL2I(Node* offset) {
1260 // short-circuit a common case
1261 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1262 if (offset_con != (jlong)Type::OffsetBot) {
1263 return intcon((int) offset_con);
1264 }
1265 return _gvn.transform( new ConvL2INode(offset));
1266 }
1267
1268 //-------------------------load_object_klass-----------------------------------
1269 Node* GraphKit::load_object_klass(Node* obj) {
1270 // Special-case a fresh allocation to avoid building nodes:
1271 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1272 if (akls != nullptr) return akls;
1273 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1274 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1275 }
1276
1277 //-------------------------load_array_length-----------------------------------
1278 Node* GraphKit::load_array_length(Node* array) {
1279 // Special-case a fresh allocation to avoid building nodes:
1280 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1281 Node *alen;
1282 if (alloc == nullptr) {
1283 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1284 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1285 } else {
1286 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1287 }
1288 return alen;
1289 }
1290
1291 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1292 const TypeOopPtr* oop_type,
1293 bool replace_length_in_map) {
1294 Node* length = alloc->Ideal_length();
1303 replace_in_map(length, ccast);
1304 }
1305 return ccast;
1306 }
1307 }
1308 return length;
1309 }
1310
1311 //------------------------------do_null_check----------------------------------
1312 // Helper function to do a null pointer check. Returned value is
1313 // the incoming address with null casted away. You are allowed to use the
1314 // not-null value only if you are control dependent on the test.
1315 #ifndef PRODUCT
1316 extern uint explicit_null_checks_inserted,
1317 explicit_null_checks_elided;
1318 #endif
1319 Node* GraphKit::null_check_common(Node* value, BasicType type,
1320 // optional arguments for variations:
1321 bool assert_null,
1322 Node* *null_control,
1323 bool speculative) {
1324 assert(!assert_null || null_control == nullptr, "not both at once");
1325 if (stopped()) return top();
1326 NOT_PRODUCT(explicit_null_checks_inserted++);
1327
1328 // Construct null check
1329 Node *chk = nullptr;
1330 switch(type) {
1331 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1332 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1333 case T_ARRAY : // fall through
1334 type = T_OBJECT; // simplify further tests
1335 case T_OBJECT : {
1336 const Type *t = _gvn.type( value );
1337
1338 const TypeOopPtr* tp = t->isa_oopptr();
1339 if (tp != nullptr && !tp->is_loaded()
1340 // Only for do_null_check, not any of its siblings:
1341 && !assert_null && null_control == nullptr) {
1342 // Usually, any field access or invocation on an unloaded oop type
1343 // will simply fail to link, since the statically linked class is
1344 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1345 // the static class is loaded but the sharper oop type is not.
1346 // Rather than checking for this obscure case in lots of places,
1347 // we simply observe that a null check on an unloaded class
1411 }
1412 Node *oldcontrol = control();
1413 set_control(cfg);
1414 Node *res = cast_not_null(value);
1415 set_control(oldcontrol);
1416 NOT_PRODUCT(explicit_null_checks_elided++);
1417 return res;
1418 }
1419 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1420 if (cfg == nullptr) break; // Quit at region nodes
1421 depth++;
1422 }
1423 }
1424
1425 //-----------
1426 // Branch to failure if null
1427 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1428 Deoptimization::DeoptReason reason;
1429 if (assert_null) {
1430 reason = Deoptimization::reason_null_assert(speculative);
1431 } else if (type == T_OBJECT) {
1432 reason = Deoptimization::reason_null_check(speculative);
1433 } else {
1434 reason = Deoptimization::Reason_div0_check;
1435 }
1436 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1437 // ciMethodData::has_trap_at will return a conservative -1 if any
1438 // must-be-null assertion has failed. This could cause performance
1439 // problems for a method after its first do_null_assert failure.
1440 // Consider using 'Reason_class_check' instead?
1441
1442 // To cause an implicit null check, we set the not-null probability
1443 // to the maximum (PROB_MAX). For an explicit check the probability
1444 // is set to a smaller value.
1445 if (null_control != nullptr || too_many_traps(reason)) {
1446 // probability is less likely
1447 ok_prob = PROB_LIKELY_MAG(3);
1448 } else if (!assert_null &&
1449 (ImplicitNullCheckThreshold > 0) &&
1450 method() != nullptr &&
1451 (method()->method_data()->trap_count(reason)
1485 }
1486
1487 if (assert_null) {
1488 // Cast obj to null on this path.
1489 replace_in_map(value, zerocon(type));
1490 return zerocon(type);
1491 }
1492
1493 // Cast obj to not-null on this path, if there is no null_control.
1494 // (If there is a null_control, a non-null value may come back to haunt us.)
1495 if (type == T_OBJECT) {
1496 Node* cast = cast_not_null(value, false);
1497 if (null_control == nullptr || (*null_control) == top())
1498 replace_in_map(value, cast);
1499 value = cast;
1500 }
1501
1502 return value;
1503 }
1504
1505
1506 //------------------------------cast_not_null----------------------------------
1507 // Cast obj to not-null on this path
1508 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1509 const Type *t = _gvn.type(obj);
1510 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1511 // Object is already not-null?
1512 if( t == t_not_null ) return obj;
1513
1514 Node* cast = new CastPPNode(control(), obj,t_not_null);
1515 cast = _gvn.transform( cast );
1516
1517 // Scan for instances of 'obj' in the current JVM mapping.
1518 // These instances are known to be not-null after the test.
1519 if (do_replace_in_map)
1520 replace_in_map(obj, cast);
1521
1522 return cast; // Return casted value
1523 }
1524
1525 // Sometimes in intrinsics, we implicitly know an object is not null
1526 // (there's no actual null check) so we can cast it to not null. In
1527 // the course of optimizations, the input to the cast can become null.
1528 // In that case that data path will die and we need the control path
1583 Node* GraphKit::memory(uint alias_idx) {
1584 MergeMemNode* mem = merged_memory();
1585 Node* p = mem->memory_at(alias_idx);
1586 assert(p != mem->empty_memory(), "empty");
1587 _gvn.set_type(p, Type::MEMORY); // must be mapped
1588 return p;
1589 }
1590
1591 //-----------------------------reset_memory------------------------------------
1592 Node* GraphKit::reset_memory() {
1593 Node* mem = map()->memory();
1594 // do not use this node for any more parsing!
1595 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1596 return _gvn.transform( mem );
1597 }
1598
1599 //------------------------------set_all_memory---------------------------------
1600 void GraphKit::set_all_memory(Node* newmem) {
1601 Node* mergemem = MergeMemNode::make(newmem);
1602 gvn().set_type_bottom(mergemem);
1603 map()->set_memory(mergemem);
1604 }
1605
1606 //------------------------------set_all_memory_call----------------------------
1607 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1608 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1609 set_all_memory(newmem);
1610 }
1611
1612 //=============================================================================
1613 //
1614 // parser factory methods for MemNodes
1615 //
1616 // These are layered on top of the factory methods in LoadNode and StoreNode,
1617 // and integrate with the parser's memory state and _gvn engine.
1618 //
1619
1620 // factory methods in "int adr_idx"
1621 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1622 MemNode::MemOrd mo,
1623 LoadNode::ControlDependency control_dependency,
1624 bool require_atomic_access,
1625 bool unaligned,
1626 bool mismatched,
1627 bool unsafe,
1628 uint8_t barrier_data) {
1629 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1630 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1631 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1632 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1633 Node* mem = memory(adr_idx);
1634 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1635 ld = _gvn.transform(ld);
1636 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1637 // Improve graph before escape analysis and boxing elimination.
1638 record_for_igvn(ld);
1639 if (ld->is_DecodeN()) {
1640 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1641 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1642 // a Phi). Recording such cases is still perfectly sound, but may be
1643 // unnecessary and result in some minor IGVN overhead.
1644 record_for_igvn(ld->in(1));
1645 }
1646 }
1647 return ld;
1648 }
1649
1650 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1651 MemNode::MemOrd mo,
1652 bool require_atomic_access,
1653 bool unaligned,
1654 bool mismatched,
1655 bool unsafe,
1669 if (unsafe) {
1670 st->as_Store()->set_unsafe_access();
1671 }
1672 st->as_Store()->set_barrier_data(barrier_data);
1673 st = _gvn.transform(st);
1674 set_memory(st, adr_idx);
1675 // Back-to-back stores can only remove intermediate store with DU info
1676 // so push on worklist for optimizer.
1677 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1678 record_for_igvn(st);
1679
1680 return st;
1681 }
1682
1683 Node* GraphKit::access_store_at(Node* obj,
1684 Node* adr,
1685 const TypePtr* adr_type,
1686 Node* val,
1687 const Type* val_type,
1688 BasicType bt,
1689 DecoratorSet decorators) {
1690 // Transformation of a value which could be null pointer (CastPP #null)
1691 // could be delayed during Parse (for example, in adjust_map_after_if()).
1692 // Execute transformation here to avoid barrier generation in such case.
1693 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1694 val = _gvn.makecon(TypePtr::NULL_PTR);
1695 }
1696
1697 if (stopped()) {
1698 return top(); // Dead path ?
1699 }
1700
1701 assert(val != nullptr, "not dead path");
1702
1703 C2AccessValuePtr addr(adr, adr_type);
1704 C2AccessValue value(val, val_type);
1705 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1706 if (access.is_raw()) {
1707 return _barrier_set->BarrierSetC2::store_at(access, value);
1708 } else {
1709 return _barrier_set->store_at(access, value);
1710 }
1711 }
1712
1713 Node* GraphKit::access_load_at(Node* obj, // containing obj
1714 Node* adr, // actual address to store val at
1715 const TypePtr* adr_type,
1716 const Type* val_type,
1717 BasicType bt,
1718 DecoratorSet decorators) {
1719 if (stopped()) {
1720 return top(); // Dead path ?
1721 }
1722
1723 SavedState old_state(this);
1724 C2AccessValuePtr addr(adr, adr_type);
1725 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1726 Node* load;
1727 if (access.is_raw()) {
1728 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1729 } else {
1730 load = _barrier_set->load_at(access, val_type);
1731 }
1732
1733 // Restore the previous state only if the load got folded to a constant
1734 // and we can discard any barriers that might have been added.
1735 if (load == nullptr || !load->is_Con()) {
1736 old_state.discard();
1737 }
1738 return load;
1739 }
1740
1741 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1742 const Type* val_type,
1743 BasicType bt,
1744 DecoratorSet decorators) {
1745 if (stopped()) {
1827 Node* new_val,
1828 const Type* value_type,
1829 BasicType bt,
1830 DecoratorSet decorators) {
1831 C2AccessValuePtr addr(adr, adr_type);
1832 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1833 if (access.is_raw()) {
1834 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1835 } else {
1836 return _barrier_set->atomic_add_at(access, new_val, value_type);
1837 }
1838 }
1839
1840 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1841 return _barrier_set->clone(this, src, dst, size, is_array);
1842 }
1843
1844 //-------------------------array_element_address-------------------------
1845 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1846 const TypeInt* sizetype, Node* ctrl) {
1847 uint shift = exact_log2(type2aelembytes(elembt));
1848 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1849
1850 // short-circuit a common case (saves lots of confusing waste motion)
1851 jint idx_con = find_int_con(idx, -1);
1852 if (idx_con >= 0) {
1853 intptr_t offset = header + ((intptr_t)idx_con << shift);
1854 return basic_plus_adr(ary, offset);
1855 }
1856
1857 // must be correct type for alignment purposes
1858 Node* base = basic_plus_adr(ary, header);
1859 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1860 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1861 return basic_plus_adr(ary, base, scale);
1862 }
1863
1864 //-------------------------load_array_element-------------------------
1865 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1866 const Type* elemtype = arytype->elem();
1867 BasicType elembt = elemtype->array_element_basic_type();
1868 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1869 if (elembt == T_NARROWOOP) {
1870 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1871 }
1872 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1873 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1874 return ld;
1875 }
1876
1877 //-------------------------set_arguments_for_java_call-------------------------
1878 // Arguments (pre-popped from the stack) are taken from the JVMS.
1879 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1880 // Add the call arguments:
1881 uint nargs = call->method()->arg_size();
1882 for (uint i = 0; i < nargs; i++) {
1883 Node* arg = argument(i);
1884 call->init_req(i + TypeFunc::Parms, arg);
1885 }
1886 }
1887
1888 //---------------------------set_edges_for_java_call---------------------------
1889 // Connect a newly created call into the current JVMS.
1890 // A return value node (if any) is returned from set_edges_for_java_call.
1891 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1892
1893 // Add the predefined inputs:
1894 call->init_req( TypeFunc::Control, control() );
1895 call->init_req( TypeFunc::I_O , i_o() );
1896 call->init_req( TypeFunc::Memory , reset_memory() );
1897 call->init_req( TypeFunc::FramePtr, frameptr() );
1898 call->init_req( TypeFunc::ReturnAdr, top() );
1899
1900 add_safepoint_edges(call, must_throw);
1901
1902 Node* xcall = _gvn.transform(call);
1903
1904 if (xcall == top()) {
1905 set_control(top());
1906 return;
1907 }
1908 assert(xcall == call, "call identity is stable");
1909
1910 // Re-use the current map to produce the result.
1911
1912 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1913 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1914 set_all_memory_call(xcall, separate_io_proj);
1915
1916 //return xcall; // no need, caller already has it
1917 }
1918
1919 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1920 if (stopped()) return top(); // maybe the call folded up?
1921
1922 // Capture the return value, if any.
1923 Node* ret;
1924 if (call->method() == nullptr ||
1925 call->method()->return_type()->basic_type() == T_VOID)
1926 ret = top();
1927 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1928
1929 // Note: Since any out-of-line call can produce an exception,
1930 // we always insert an I_O projection from the call into the result.
1931
1932 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1933
1934 if (separate_io_proj) {
1935 // The caller requested separate projections be used by the fall
1936 // through and exceptional paths, so replace the projections for
1937 // the fall through path.
1938 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1939 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1940 }
1941 return ret;
1942 }
1943
1944 //--------------------set_predefined_input_for_runtime_call--------------------
1945 // Reading and setting the memory state is way conservative here.
1946 // The real problem is that I am not doing real Type analysis on memory,
1947 // so I cannot distinguish card mark stores from other stores. Across a GC
1948 // point the Store Barrier and the card mark memory has to agree. I cannot
1949 // have a card mark store and its barrier split across the GC point from
1950 // either above or below. Here I get that to happen by reading ALL of memory.
1951 // A better answer would be to separate out card marks from other memory.
1952 // For now, return the input memory state, so that it can be reused
1953 // after the call, if this call has restricted memory effects.
1954 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1955 // Set fixed predefined input arguments
1956 call->init_req(TypeFunc::Control, control());
1957 call->init_req(TypeFunc::I_O, top()); // does no i/o
1958 call->init_req(TypeFunc::ReturnAdr, top());
1959 if (call->is_CallLeafPure()) {
1960 call->init_req(TypeFunc::Memory, top());
2022 if (use->is_MergeMem()) {
2023 wl.push(use);
2024 }
2025 }
2026 }
2027
2028 // Replace the call with the current state of the kit.
2029 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2030 JVMState* ejvms = nullptr;
2031 if (has_exceptions()) {
2032 ejvms = transfer_exceptions_into_jvms();
2033 }
2034
2035 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2036 ReplacedNodes replaced_nodes_exception;
2037 Node* ex_ctl = top();
2038
2039 SafePointNode* final_state = stop();
2040
2041 // Find all the needed outputs of this call
2042 CallProjections callprojs;
2043 call->extract_projections(&callprojs, true, do_asserts);
2044
2045 Unique_Node_List wl;
2046 Node* init_mem = call->in(TypeFunc::Memory);
2047 Node* final_mem = final_state->in(TypeFunc::Memory);
2048 Node* final_ctl = final_state->in(TypeFunc::Control);
2049 Node* final_io = final_state->in(TypeFunc::I_O);
2050
2051 // Replace all the old call edges with the edges from the inlining result
2052 if (callprojs.fallthrough_catchproj != nullptr) {
2053 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
2054 }
2055 if (callprojs.fallthrough_memproj != nullptr) {
2056 if (final_mem->is_MergeMem()) {
2057 // Parser's exits MergeMem was not transformed but may be optimized
2058 final_mem = _gvn.transform(final_mem);
2059 }
2060 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
2061 add_mergemem_users_to_worklist(wl, final_mem);
2062 }
2063 if (callprojs.fallthrough_ioproj != nullptr) {
2064 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
2065 }
2066
2067 // Replace the result with the new result if it exists and is used
2068 if (callprojs.resproj != nullptr && result != nullptr) {
2069 C->gvn_replace_by(callprojs.resproj, result);
2070 }
2071
2072 if (ejvms == nullptr) {
2073 // No exception edges to simply kill off those paths
2074 if (callprojs.catchall_catchproj != nullptr) {
2075 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2076 }
2077 if (callprojs.catchall_memproj != nullptr) {
2078 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
2079 }
2080 if (callprojs.catchall_ioproj != nullptr) {
2081 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
2082 }
2083 // Replace the old exception object with top
2084 if (callprojs.exobj != nullptr) {
2085 C->gvn_replace_by(callprojs.exobj, C->top());
2086 }
2087 } else {
2088 GraphKit ekit(ejvms);
2089
2090 // Load my combined exception state into the kit, with all phis transformed:
2091 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2092 replaced_nodes_exception = ex_map->replaced_nodes();
2093
2094 Node* ex_oop = ekit.use_exception_state(ex_map);
2095
2096 if (callprojs.catchall_catchproj != nullptr) {
2097 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2098 ex_ctl = ekit.control();
2099 }
2100 if (callprojs.catchall_memproj != nullptr) {
2101 Node* ex_mem = ekit.reset_memory();
2102 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2103 add_mergemem_users_to_worklist(wl, ex_mem);
2104 }
2105 if (callprojs.catchall_ioproj != nullptr) {
2106 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2107 }
2108
2109 // Replace the old exception object with the newly created one
2110 if (callprojs.exobj != nullptr) {
2111 C->gvn_replace_by(callprojs.exobj, ex_oop);
2112 }
2113 }
2114
2115 // Disconnect the call from the graph
2116 call->disconnect_inputs(C);
2117 C->gvn_replace_by(call, C->top());
2118
2119 // Clean up any MergeMems that feed other MergeMems since the
2120 // optimizer doesn't like that.
2121 while (wl.size() > 0) {
2122 _gvn.transform(wl.pop());
2123 }
2124
2125 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2126 replaced_nodes.apply(C, final_ctl);
2127 }
2128 if (!ex_ctl->is_top() && do_replaced_nodes) {
2129 replaced_nodes_exception.apply(C, ex_ctl);
2130 }
2131 }
2132
2133
2134 //------------------------------increment_counter------------------------------
2135 // for statistics: increment a VM counter by 1
2136
2137 void GraphKit::increment_counter(address counter_addr) {
2138 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2139 increment_counter(adr1);
2140 }
2141
2142 void GraphKit::increment_counter(Node* counter_addr) {
2143 Node* ctrl = control();
2144 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2145 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2146 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2147 }
2148
2149 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2150 Node* halt = new HaltNode(ctrl, frameptr, reason
2151 PRODUCT_ONLY(COMMA generate_code_in_product));
2152 halt = _gvn.transform(halt);
2153 root()->add_req(halt);
2154 }
2155
2156 //------------------------------uncommon_trap----------------------------------
2157 // Bail out to the interpreter in mid-method. Implemented by calling the
2158 // uncommon_trap blob. This helper function inserts a runtime call with the
2159 // right debug info.
2160 Node* GraphKit::uncommon_trap(int trap_request,
2161 ciKlass* klass, const char* comment,
2162 bool must_throw,
2163 bool keep_exact_action) {
2164 if (failing_internal()) {
2165 stop();
2166 }
2167 if (stopped()) return nullptr; // trap reachable?
2168
2169 // Note: If ProfileTraps is true, and if a deopt. actually
2170 // occurs here, the runtime will make sure an MDO exists. There is
2171 // no need to call method()->ensure_method_data() at this point.
2172
2173 // Set the stack pointer to the right value for reexecution:
2315 *
2316 * @param n node that the type applies to
2317 * @param exact_kls type from profiling
2318 * @param maybe_null did profiling see null?
2319 *
2320 * @return node with improved type
2321 */
2322 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2323 const Type* current_type = _gvn.type(n);
2324 assert(UseTypeSpeculation, "type speculation must be on");
2325
2326 const TypePtr* speculative = current_type->speculative();
2327
2328 // Should the klass from the profile be recorded in the speculative type?
2329 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2330 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2331 const TypeOopPtr* xtype = tklass->as_instance_type();
2332 assert(xtype->klass_is_exact(), "Should be exact");
2333 // Any reason to believe n is not null (from this profiling or a previous one)?
2334 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2335 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2336 // record the new speculative type's depth
2337 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2338 speculative = speculative->with_inline_depth(jvms()->depth());
2339 } else if (current_type->would_improve_ptr(ptr_kind)) {
2340 // Profiling report that null was never seen so we can change the
2341 // speculative type to non null ptr.
2342 if (ptr_kind == ProfileAlwaysNull) {
2343 speculative = TypePtr::NULL_PTR;
2344 } else {
2345 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2346 const TypePtr* ptr = TypePtr::NOTNULL;
2347 if (speculative != nullptr) {
2348 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2349 } else {
2350 speculative = ptr;
2351 }
2352 }
2353 }
2354
2355 if (speculative != current_type->speculative()) {
2356 // Build a type with a speculative type (what we think we know
2357 // about the type but will need a guard when we use it)
2358 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2359 // We're changing the type, we need a new CheckCast node to carry
2360 // the new type. The new type depends on the control: what
2361 // profiling tells us is only valid from here as far as we can
2362 // tell.
2363 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2364 cast = _gvn.transform(cast);
2365 replace_in_map(n, cast);
2366 n = cast;
2367 }
2368
2369 return n;
2370 }
2371
2372 /**
2373 * Record profiling data from receiver profiling at an invoke with the
2374 * type system so that it can propagate it (speculation)
2375 *
2376 * @param n receiver node
2377 *
2378 * @return node with improved type
2379 */
2380 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2381 if (!UseTypeSpeculation) {
2382 return n;
2383 }
2384 ciKlass* exact_kls = profile_has_unique_klass();
2385 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2386 if ((java_bc() == Bytecodes::_checkcast ||
2387 java_bc() == Bytecodes::_instanceof ||
2388 java_bc() == Bytecodes::_aastore) &&
2389 method()->method_data()->is_mature()) {
2390 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2391 if (data != nullptr) {
2392 if (!data->as_BitData()->null_seen()) {
2393 ptr_kind = ProfileNeverNull;
2394 } else {
2395 if (TypeProfileCasts) {
2396 assert(data->is_ReceiverTypeData(), "bad profile data type");
2397 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2398 uint i = 0;
2399 for (; i < call->row_limit(); i++) {
2400 ciKlass* receiver = call->receiver(i);
2401 if (receiver != nullptr) {
2402 break;
2403 }
2404 }
2405 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2406 }
2407 }
2408 }
2409 }
2410 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2411 }
2412
2413 /**
2414 * Record profiling data from argument profiling at an invoke with the
2415 * type system so that it can propagate it (speculation)
2416 *
2417 * @param dest_method target method for the call
2418 * @param bc what invoke bytecode is this?
2419 */
2420 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2421 if (!UseTypeSpeculation) {
2422 return;
2423 }
2424 const TypeFunc* tf = TypeFunc::make(dest_method);
2425 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2426 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2427 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2428 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2429 if (is_reference_type(targ->basic_type())) {
2430 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2431 ciKlass* better_type = nullptr;
2432 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2433 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2434 }
2435 i++;
2436 }
2437 }
2438 }
2439
2440 /**
2441 * Record profiling data from parameter profiling at an invoke with
2442 * the type system so that it can propagate it (speculation)
2443 */
2444 void GraphKit::record_profiled_parameters_for_speculation() {
2445 if (!UseTypeSpeculation) {
2446 return;
2447 }
2448 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2568 // The first null ends the list.
2569 Node* parm0, Node* parm1,
2570 Node* parm2, Node* parm3,
2571 Node* parm4, Node* parm5,
2572 Node* parm6, Node* parm7) {
2573 assert(call_addr != nullptr, "must not call null targets");
2574
2575 // Slow-path call
2576 bool is_leaf = !(flags & RC_NO_LEAF);
2577 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2578 if (call_name == nullptr) {
2579 assert(!is_leaf, "must supply name for leaf");
2580 call_name = OptoRuntime::stub_name(call_addr);
2581 }
2582 CallNode* call;
2583 if (!is_leaf) {
2584 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2585 } else if (flags & RC_NO_FP) {
2586 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2587 } else if (flags & RC_VECTOR){
2588 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2589 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2590 } else if (flags & RC_PURE) {
2591 assert(adr_type == nullptr, "pure call does not touch memory");
2592 call = new CallLeafPureNode(call_type, call_addr, call_name);
2593 } else {
2594 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2595 }
2596
2597 // The following is similar to set_edges_for_java_call,
2598 // except that the memory effects of the call are restricted to AliasIdxRaw.
2599
2600 // Slow path call has no side-effects, uses few values
2601 bool wide_in = !(flags & RC_NARROW_MEM);
2602 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2603
2604 Node* prev_mem = nullptr;
2605 if (wide_in) {
2606 prev_mem = set_predefined_input_for_runtime_call(call);
2607 } else {
2608 assert(!wide_out, "narrow in => narrow out");
2609 Node* narrow_mem = memory(adr_type);
2610 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2611 }
2612
2613 // Hook each parm in order. Stop looking at the first null.
2614 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2615 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2616 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2617 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2618 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2619 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2620 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2621 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2622 /* close each nested if ===> */ } } } } } } } }
2623 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2624
2625 if (!is_leaf) {
2626 // Non-leaves can block and take safepoints:
2627 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2628 }
2629 // Non-leaves can throw exceptions:
2630 if (has_io) {
2631 call->set_req(TypeFunc::I_O, i_o());
2632 }
2633
2634 if (flags & RC_UNCOMMON) {
2635 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2636 // (An "if" probability corresponds roughly to an unconditional count.
2637 // Sort of.)
2638 call->set_cnt(PROB_UNLIKELY_MAG(4));
2639 }
2640
2641 Node* c = _gvn.transform(call);
2642 assert(c == call, "cannot disappear");
2643
2651
2652 if (has_io) {
2653 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2654 }
2655 return call;
2656
2657 }
2658
2659 // i2b
2660 Node* GraphKit::sign_extend_byte(Node* in) {
2661 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2662 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2663 }
2664
2665 // i2s
2666 Node* GraphKit::sign_extend_short(Node* in) {
2667 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2668 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2669 }
2670
2671 //------------------------------merge_memory-----------------------------------
2672 // Merge memory from one path into the current memory state.
2673 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2674 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2675 Node* old_slice = mms.force_memory();
2676 Node* new_slice = mms.memory2();
2677 if (old_slice != new_slice) {
2678 PhiNode* phi;
2679 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2680 if (mms.is_empty()) {
2681 // clone base memory Phi's inputs for this memory slice
2682 assert(old_slice == mms.base_memory(), "sanity");
2683 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2684 _gvn.set_type(phi, Type::MEMORY);
2685 for (uint i = 1; i < phi->req(); i++) {
2686 phi->init_req(i, old_slice->in(i));
2687 }
2688 } else {
2689 phi = old_slice->as_Phi(); // Phi was generated already
2690 }
2747 gvn.transform(iff);
2748 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2749 return iff;
2750 }
2751
2752 //-------------------------------gen_subtype_check-----------------------------
2753 // Generate a subtyping check. Takes as input the subtype and supertype.
2754 // Returns 2 values: sets the default control() to the true path and returns
2755 // the false path. Only reads invariant memory; sets no (visible) memory.
2756 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2757 // but that's not exposed to the optimizer. This call also doesn't take in an
2758 // Object; if you wish to check an Object you need to load the Object's class
2759 // prior to coming here.
2760 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2761 ciMethod* method, int bci) {
2762 Compile* C = gvn.C;
2763 if ((*ctrl)->is_top()) {
2764 return C->top();
2765 }
2766
2767 // Fast check for identical types, perhaps identical constants.
2768 // The types can even be identical non-constants, in cases
2769 // involving Array.newInstance, Object.clone, etc.
2770 if (subklass == superklass)
2771 return C->top(); // false path is dead; no test needed.
2772
2773 if (gvn.type(superklass)->singleton()) {
2774 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2775 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2776
2777 // In the common case of an exact superklass, try to fold up the
2778 // test before generating code. You may ask, why not just generate
2779 // the code and then let it fold up? The answer is that the generated
2780 // code will necessarily include null checks, which do not always
2781 // completely fold away. If they are also needless, then they turn
2782 // into a performance loss. Example:
2783 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2784 // Here, the type of 'fa' is often exact, so the store check
2785 // of fa[1]=x will fold up, without testing the nullness of x.
2786 //
2787 // At macro expansion, we would have already folded the SubTypeCheckNode
2788 // being expanded here because we always perform the static sub type
2789 // check in SubTypeCheckNode::sub() regardless of whether
2790 // StressReflectiveCode is set or not. We can therefore skip this
2791 // static check when StressReflectiveCode is on.
2792 switch (C->static_subtype_check(superk, subk)) {
2793 case Compile::SSC_always_false:
2794 {
2795 Node* always_fail = *ctrl;
2796 *ctrl = gvn.C->top();
2797 return always_fail;
2798 }
2799 case Compile::SSC_always_true:
2800 return C->top();
2801 case Compile::SSC_easy_test:
2802 {
2803 // Just do a direct pointer compare and be done.
2804 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2805 *ctrl = gvn.transform(new IfTrueNode(iff));
2806 return gvn.transform(new IfFalseNode(iff));
2807 }
2808 case Compile::SSC_full_test:
2809 break;
2810 default:
2811 ShouldNotReachHere();
2812 }
2813 }
2814
2815 // %%% Possible further optimization: Even if the superklass is not exact,
2816 // if the subklass is the unique subtype of the superklass, the check
2817 // will always succeed. We could leave a dependency behind to ensure this.
2818
2819 // First load the super-klass's check-offset
2820 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2821 Node* m = C->immutable_memory();
2822 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2823 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2824 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2862 gvn.record_for_igvn(r_ok_subtype);
2863
2864 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2865 // SubTypeCheck node
2866 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2867 ciCallProfile profile = method->call_profile_at_bci(bci);
2868 float total_prob = 0;
2869 for (int i = 0; profile.has_receiver(i); ++i) {
2870 float prob = profile.receiver_prob(i);
2871 total_prob += prob;
2872 }
2873 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2874 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2875 for (int i = 0; profile.has_receiver(i); ++i) {
2876 ciKlass* klass = profile.receiver(i);
2877 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2878 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2879 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2880 continue;
2881 }
2882 float prob = profile.receiver_prob(i);
2883 ConNode* klass_node = gvn.makecon(klass_t);
2884 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2885 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2886
2887 if (result == Compile::SSC_always_true) {
2888 r_ok_subtype->add_req(iftrue);
2889 } else {
2890 assert(result == Compile::SSC_always_false, "");
2891 r_not_subtype->add_req(iftrue);
2892 }
2893 *ctrl = gvn.transform(new IfFalseNode(iff));
2894 }
2895 }
2896 }
2897
2898 // See if we get an immediate positive hit. Happens roughly 83% of the
2899 // time. Test to see if the value loaded just previously from the subklass
2900 // is exactly the superklass.
2901 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2915 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
2916 }
2917 return not_subtype_ctrl;
2918 }
2919
2920 r_ok_subtype->init_req(1, iftrue1);
2921
2922 // Check for immediate negative hit. Happens roughly 11% of the time (which
2923 // is roughly 63% of the remaining cases). Test to see if the loaded
2924 // check-offset points into the subklass display list or the 1-element
2925 // cache. If it points to the display (and NOT the cache) and the display
2926 // missed then it's not a subtype.
2927 Node *cacheoff = gvn.intcon(cacheoff_con);
2928 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2929 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2930 *ctrl = gvn.transform(new IfFalseNode(iff2));
2931
2932 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2933 // No performance impact (too rare) but allows sharing of secondary arrays
2934 // which has some footprint reduction.
2935 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2936 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2937 *ctrl = gvn.transform(new IfFalseNode(iff3));
2938
2939 // -- Roads not taken here: --
2940 // We could also have chosen to perform the self-check at the beginning
2941 // of this code sequence, as the assembler does. This would not pay off
2942 // the same way, since the optimizer, unlike the assembler, can perform
2943 // static type analysis to fold away many successful self-checks.
2944 // Non-foldable self checks work better here in second position, because
2945 // the initial primary superclass check subsumes a self-check for most
2946 // types. An exception would be a secondary type like array-of-interface,
2947 // which does not appear in its own primary supertype display.
2948 // Finally, we could have chosen to move the self-check into the
2949 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2950 // dependent manner. But it is worthwhile to have the check here,
2951 // where it can be perhaps be optimized. The cost in code space is
2952 // small (register compare, branch).
2953
2954 // Now do a linear scan of the secondary super-klass array. Again, no real
2955 // performance impact (too rare) but it's gotta be done.
2956 // Since the code is rarely used, there is no penalty for moving it
2957 // out of line, and it can only improve I-cache density.
2958 // The decision to inline or out-of-line this final check is platform
2959 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2960 Node* psc = gvn.transform(
2961 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2962
2963 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2964 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2965 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2966
2967 // Return false path; set default control to true path.
2968 *ctrl = gvn.transform(r_ok_subtype);
2969 return gvn.transform(r_not_subtype);
2970 }
2971
2972 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2973 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2974 if (expand_subtype_check) {
2975 MergeMemNode* mem = merged_memory();
2976 Node* ctrl = control();
2977 Node* subklass = obj_or_subklass;
2978 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2979 subklass = load_object_klass(obj_or_subklass);
2980 }
2981
2982 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2983 set_control(ctrl);
2984 return n;
2985 }
2986
2987 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2988 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2989 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2990 set_control(_gvn.transform(new IfTrueNode(iff)));
2991 return _gvn.transform(new IfFalseNode(iff));
2992 }
2993
2994 // Profile-driven exact type check:
2995 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2996 float prob,
2997 Node* *casted_receiver) {
2998 assert(!klass->is_interface(), "no exact type check on interfaces");
2999
3000 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3001 Node* recv_klass = load_object_klass(receiver);
3002 Node* want_klass = makecon(tklass);
3003 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3004 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3005 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3006 set_control( _gvn.transform(new IfTrueNode (iff)));
3007 Node* fail = _gvn.transform(new IfFalseNode(iff));
3008
3009 if (!stopped()) {
3010 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3011 const TypeOopPtr* recvx_type = tklass->as_instance_type();
3012 assert(recvx_type->klass_is_exact(), "");
3013
3014 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
3015 // Subsume downstream occurrences of receiver with a cast to
3016 // recv_xtype, since now we know what the type will be.
3017 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
3018 (*casted_receiver) = _gvn.transform(cast);
3019 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3020 // (User must make the replace_in_map call.)
3021 }
3022 }
3023
3024 return fail;
3025 }
3026
3027 //------------------------------subtype_check_receiver-------------------------
3028 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3029 Node** casted_receiver) {
3030 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3031 Node* want_klass = makecon(tklass);
3032
3033 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3034
3035 // Ignore interface type information until interface types are properly tracked.
3036 if (!stopped() && !klass->is_interface()) {
3037 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3038 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3039 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3040 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3041 (*casted_receiver) = _gvn.transform(cast);
3042 }
3043 }
3044
3045 return slow_ctl;
3046 }
3047
3048 //------------------------------seems_never_null-------------------------------
3049 // Use null_seen information if it is available from the profile.
3050 // If we see an unexpected null at a type check we record it and force a
3051 // recompile; the offending check will be recompiled to handle nulls.
3052 // If we see several offending BCIs, then all checks in the
3053 // method will be recompiled.
3054 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3055 speculating = !_gvn.type(obj)->speculative_maybe_null();
3056 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3057 if (UncommonNullCast // Cutout for this technique
3058 && obj != null() // And not the -Xcomp stupid case?
3059 && !too_many_traps(reason)
3060 ) {
3061 if (speculating) {
3130
3131 //------------------------maybe_cast_profiled_receiver-------------------------
3132 // If the profile has seen exactly one type, narrow to exactly that type.
3133 // Subsequent type checks will always fold up.
3134 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3135 const TypeKlassPtr* require_klass,
3136 ciKlass* spec_klass,
3137 bool safe_for_replace) {
3138 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3139
3140 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3141
3142 // Make sure we haven't already deoptimized from this tactic.
3143 if (too_many_traps_or_recompiles(reason))
3144 return nullptr;
3145
3146 // (No, this isn't a call, but it's enough like a virtual call
3147 // to use the same ciMethod accessor to get the profile info...)
3148 // If we have a speculative type use it instead of profiling (which
3149 // may not help us)
3150 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3151 if (exact_kls != nullptr) {// no cast failures here
3152 if (require_klass == nullptr ||
3153 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3154 // If we narrow the type to match what the type profile sees or
3155 // the speculative type, we can then remove the rest of the
3156 // cast.
3157 // This is a win, even if the exact_kls is very specific,
3158 // because downstream operations, such as method calls,
3159 // will often benefit from the sharper type.
3160 Node* exact_obj = not_null_obj; // will get updated in place...
3161 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3162 &exact_obj);
3163 { PreserveJVMState pjvms(this);
3164 set_control(slow_ctl);
3165 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3166 }
3167 if (safe_for_replace) {
3168 replace_in_map(not_null_obj, exact_obj);
3169 }
3170 return exact_obj;
3260 // If not_null_obj is dead, only null-path is taken
3261 if (stopped()) { // Doing instance-of on a null?
3262 set_control(null_ctl);
3263 return intcon(0);
3264 }
3265 region->init_req(_null_path, null_ctl);
3266 phi ->init_req(_null_path, intcon(0)); // Set null path value
3267 if (null_ctl == top()) {
3268 // Do this eagerly, so that pattern matches like is_diamond_phi
3269 // will work even during parsing.
3270 assert(_null_path == PATH_LIMIT-1, "delete last");
3271 region->del_req(_null_path);
3272 phi ->del_req(_null_path);
3273 }
3274
3275 // Do we know the type check always succeed?
3276 bool known_statically = false;
3277 if (_gvn.type(superklass)->singleton()) {
3278 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3279 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3280 if (subk->is_loaded()) {
3281 int static_res = C->static_subtype_check(superk, subk);
3282 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3283 }
3284 }
3285
3286 if (!known_statically) {
3287 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3288 // We may not have profiling here or it may not help us. If we
3289 // have a speculative type use it to perform an exact cast.
3290 ciKlass* spec_obj_type = obj_type->speculative_type();
3291 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3292 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3293 if (stopped()) { // Profile disagrees with this path.
3294 set_control(null_ctl); // Null is the only remaining possibility.
3295 return intcon(0);
3296 }
3297 if (cast_obj != nullptr) {
3298 not_null_obj = cast_obj;
3299 }
3300 }
3316 record_for_igvn(region);
3317
3318 // If we know the type check always succeeds then we don't use the
3319 // profiling data at this bytecode. Don't lose it, feed it to the
3320 // type system as a speculative type.
3321 if (safe_for_replace) {
3322 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3323 replace_in_map(obj, casted_obj);
3324 }
3325
3326 return _gvn.transform(phi);
3327 }
3328
3329 //-------------------------------gen_checkcast---------------------------------
3330 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3331 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3332 // uncommon-trap paths work. Adjust stack after this call.
3333 // If failure_control is supplied and not null, it is filled in with
3334 // the control edge for the cast failure. Otherwise, an appropriate
3335 // uncommon trap or exception is thrown.
3336 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3337 Node* *failure_control) {
3338 kill_dead_locals(); // Benefit all the uncommon traps
3339 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3340 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3341 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3342
3343 // Fast cutout: Check the case that the cast is vacuously true.
3344 // This detects the common cases where the test will short-circuit
3345 // away completely. We do this before we perform the null check,
3346 // because if the test is going to turn into zero code, we don't
3347 // want a residual null check left around. (Causes a slowdown,
3348 // for example, in some objArray manipulations, such as a[i]=a[j].)
3349 if (improved_klass_ptr_type->singleton()) {
3350 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3351 if (objtp != nullptr) {
3352 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3353 case Compile::SSC_always_true:
3354 // If we know the type check always succeed then we don't use
3355 // the profiling data at this bytecode. Don't lose it, feed it
3356 // to the type system as a speculative type.
3357 return record_profiled_receiver_for_speculation(obj);
3358 case Compile::SSC_always_false:
3359 // It needs a null check because a null will *pass* the cast check.
3360 // A non-null value will always produce an exception.
3361 if (!objtp->maybe_null()) {
3362 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3363 Deoptimization::DeoptReason reason = is_aastore ?
3364 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3365 builtin_throw(reason);
3366 return top();
3367 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3368 return null_assert(obj);
3369 }
3370 break; // Fall through to full check
3371 default:
3372 break;
3373 }
3374 }
3375 }
3376
3377 ciProfileData* data = nullptr;
3378 bool safe_for_replace = false;
3379 if (failure_control == nullptr) { // use MDO in regular case only
3380 assert(java_bc() == Bytecodes::_aastore ||
3381 java_bc() == Bytecodes::_checkcast,
3382 "interpreter profiles type checks only for these BCs");
3383 data = method()->method_data()->bci_to_data(bci());
3384 safe_for_replace = true;
3385 }
3386
3387 // Make the merge point
3388 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3389 RegionNode* region = new RegionNode(PATH_LIMIT);
3390 Node* phi = new PhiNode(region, toop);
3391 C->set_has_split_ifs(true); // Has chance for split-if optimization
3392
3393 // Use null-cast information if it is available
3394 bool speculative_not_null = false;
3395 bool never_see_null = ((failure_control == nullptr) // regular case only
3396 && seems_never_null(obj, data, speculative_not_null));
3397
3398 // Null check; get casted pointer; set region slot 3
3399 Node* null_ctl = top();
3400 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3401
3402 // If not_null_obj is dead, only null-path is taken
3403 if (stopped()) { // Doing instance-of on a null?
3404 set_control(null_ctl);
3405 return null();
3406 }
3407 region->init_req(_null_path, null_ctl);
3408 phi ->init_req(_null_path, null()); // Set null path value
3409 if (null_ctl == top()) {
3410 // Do this eagerly, so that pattern matches like is_diamond_phi
3411 // will work even during parsing.
3412 assert(_null_path == PATH_LIMIT-1, "delete last");
3413 region->del_req(_null_path);
3414 phi ->del_req(_null_path);
3415 }
3416
3417 Node* cast_obj = nullptr;
3418 if (improved_klass_ptr_type->klass_is_exact()) {
3419 // The following optimization tries to statically cast the speculative type of the object
3420 // (for example obtained during profiling) to the type of the superklass and then do a
3421 // dynamic check that the type of the object is what we expect. To work correctly
3422 // for checkcast and aastore the type of superklass should be exact.
3423 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3424 // We may not have profiling here or it may not help us. If we have
3425 // a speculative type use it to perform an exact cast.
3426 ciKlass* spec_obj_type = obj_type->speculative_type();
3427 if (spec_obj_type != nullptr || data != nullptr) {
3428 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3429 if (cast_obj != nullptr) {
3430 if (failure_control != nullptr) // failure is now impossible
3431 (*failure_control) = top();
3432 // adjust the type of the phi to the exact klass:
3433 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3434 }
3435 }
3436 }
3437
3438 if (cast_obj == nullptr) {
3439 // Generate the subtype check
3440 Node* improved_superklass = superklass;
3441 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3442 improved_superklass = makecon(improved_klass_ptr_type);
3443 }
3444 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3445
3446 // Plug in success path into the merge
3447 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3448 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3449 if (failure_control == nullptr) {
3450 if (not_subtype_ctrl != top()) { // If failure is possible
3451 PreserveJVMState pjvms(this);
3452 set_control(not_subtype_ctrl);
3453 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3454 Deoptimization::DeoptReason reason = is_aastore ?
3455 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3456 builtin_throw(reason);
3457 }
3458 } else {
3459 (*failure_control) = not_subtype_ctrl;
3460 }
3461 }
3462
3463 region->init_req(_obj_path, control());
3464 phi ->init_req(_obj_path, cast_obj);
3465
3466 // A merge of null or Casted-NotNull obj
3467 Node* res = _gvn.transform(phi);
3468
3469 // Note I do NOT always 'replace_in_map(obj,result)' here.
3470 // if( tk->klass()->can_be_primary_super() )
3471 // This means that if I successfully store an Object into an array-of-String
3472 // I 'forget' that the Object is really now known to be a String. I have to
3473 // do this because we don't have true union types for interfaces - if I store
3474 // a Baz into an array-of-Interface and then tell the optimizer it's an
3475 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3476 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3477 // replace_in_map( obj, res );
3478
3479 // Return final merged results
3480 set_control( _gvn.transform(region) );
3481 record_for_igvn(region);
3482
3483 return record_profiled_receiver_for_speculation(res);
3484 }
3485
3486 //------------------------------next_monitor-----------------------------------
3487 // What number should be given to the next monitor?
3488 int GraphKit::next_monitor() {
3489 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3490 int next = current + C->sync_stack_slots();
3491 // Keep the toplevel high water mark current:
3492 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3493 return current;
3494 }
3495
3496 //------------------------------insert_mem_bar---------------------------------
3497 // Memory barrier to avoid floating things around
3498 // The membar serves as a pinch point between both control and all memory slices.
3499 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3500 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3501 mb->init_req(TypeFunc::Control, control());
3502 mb->init_req(TypeFunc::Memory, reset_memory());
3503 Node* membar = _gvn.transform(mb);
3597 lock->create_lock_counter(map()->jvms());
3598 increment_counter(lock->counter()->addr());
3599 }
3600 #endif
3601
3602 return flock;
3603 }
3604
3605
3606 //------------------------------shared_unlock----------------------------------
3607 // Emit unlocking code.
3608 void GraphKit::shared_unlock(Node* box, Node* obj) {
3609 // bci is either a monitorenter bc or InvocationEntryBci
3610 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3611 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3612
3613 if (stopped()) { // Dead monitor?
3614 map()->pop_monitor(); // Kill monitor from debug info
3615 return;
3616 }
3617
3618 // Memory barrier to avoid floating things down past the locked region
3619 insert_mem_bar(Op_MemBarReleaseLock);
3620
3621 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3622 UnlockNode *unlock = new UnlockNode(C, tf);
3623 #ifdef ASSERT
3624 unlock->set_dbg_jvms(sync_jvms());
3625 #endif
3626 uint raw_idx = Compile::AliasIdxRaw;
3627 unlock->init_req( TypeFunc::Control, control() );
3628 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3629 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3630 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3631 unlock->init_req( TypeFunc::ReturnAdr, top() );
3632
3633 unlock->init_req(TypeFunc::Parms + 0, obj);
3634 unlock->init_req(TypeFunc::Parms + 1, box);
3635 unlock = _gvn.transform(unlock)->as_Unlock();
3636
3637 Node* mem = reset_memory();
3638
3639 // unlock has no side-effects, sets few values
3640 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3641
3642 // Kill monitor from debug info
3643 map()->pop_monitor( );
3644 }
3645
3646 //-------------------------------get_layout_helper-----------------------------
3647 // If the given klass is a constant or known to be an array,
3648 // fetch the constant layout helper value into constant_value
3649 // and return null. Otherwise, load the non-constant
3650 // layout helper value, and return the node which represents it.
3651 // This two-faced routine is useful because allocation sites
3652 // almost always feature constant types.
3653 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3654 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3655 if (!StressReflectiveCode && klass_t != nullptr) {
3656 bool xklass = klass_t->klass_is_exact();
3657 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3658 jint lhelper;
3659 if (klass_t->isa_aryklassptr()) {
3660 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3661 if (is_reference_type(elem, true)) {
3662 elem = T_OBJECT;
3663 }
3664 lhelper = Klass::array_layout_helper(elem);
3665 } else {
3666 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3667 }
3668 if (lhelper != Klass::_lh_neutral_value) {
3669 constant_value = lhelper;
3670 return (Node*) nullptr;
3671 }
3672 }
3673 }
3674 constant_value = Klass::_lh_neutral_value; // put in a known value
3675 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
3676 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3677 }
3678
3679 // We just put in an allocate/initialize with a big raw-memory effect.
3680 // Hook selected additional alias categories on the initialization.
3681 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3682 MergeMemNode* init_in_merge,
3683 Node* init_out_raw) {
3684 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3685 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3686
3687 Node* prevmem = kit.memory(alias_idx);
3688 init_in_merge->set_memory_at(alias_idx, prevmem);
3689 kit.set_memory(init_out_raw, alias_idx);
3690 }
3691
3692 //---------------------------set_output_for_allocation-------------------------
3693 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3694 const TypeOopPtr* oop_type,
3695 bool deoptimize_on_exception) {
3696 int rawidx = Compile::AliasIdxRaw;
3697 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3698 add_safepoint_edges(alloc);
3699 Node* allocx = _gvn.transform(alloc);
3700 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3701 // create memory projection for i_o
3702 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3703 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3704
3705 // create a memory projection as for the normal control path
3706 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3707 set_memory(malloc, rawidx);
3708
3709 // a normal slow-call doesn't change i_o, but an allocation does
3710 // we create a separate i_o projection for the normal control path
3711 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3712 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3713
3714 // put in an initialization barrier
3715 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3716 rawoop)->as_Initialize();
3717 assert(alloc->initialization() == init, "2-way macro link must work");
3718 assert(init ->allocation() == alloc, "2-way macro link must work");
3719 {
3720 // Extract memory strands which may participate in the new object's
3721 // initialization, and source them from the new InitializeNode.
3722 // This will allow us to observe initializations when they occur,
3723 // and link them properly (as a group) to the InitializeNode.
3724 assert(init->in(InitializeNode::Memory) == malloc, "");
3725 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3726 init->set_req(InitializeNode::Memory, minit_in);
3727 record_for_igvn(minit_in); // fold it up later, if possible
3728 Node* minit_out = memory(rawidx);
3729 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3730 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
3731 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
3732 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
3733 // multiple projections as a result.
3734 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
3735 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
3736 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
3737 if (oop_type->isa_aryptr()) {
3738 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3739 int elemidx = C->get_alias_index(telemref);
3740 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));
3741 } else if (oop_type->isa_instptr()) {
3742 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3743 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3744 ciField* field = ik->nonstatic_field_at(i);
3745 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3746 continue; // do not bother to track really large numbers of fields
3747 // Find (or create) the alias category for this field:
3748 int fieldidx = C->alias_type(field)->index();
3749 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
3750 }
3751 }
3752 }
3753
3754 // Cast raw oop to the real thing...
3755 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3756 javaoop = _gvn.transform(javaoop);
3757 C->set_recent_alloc(control(), javaoop);
3758 assert(just_allocated_object(control()) == javaoop, "just allocated");
3759
3760 #ifdef ASSERT
3772 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3773 }
3774 }
3775 #endif //ASSERT
3776
3777 return javaoop;
3778 }
3779
3780 //---------------------------new_instance--------------------------------------
3781 // This routine takes a klass_node which may be constant (for a static type)
3782 // or may be non-constant (for reflective code). It will work equally well
3783 // for either, and the graph will fold nicely if the optimizer later reduces
3784 // the type to a constant.
3785 // The optional arguments are for specialized use by intrinsics:
3786 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3787 // - If 'return_size_val', report the total object size to the caller.
3788 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3789 Node* GraphKit::new_instance(Node* klass_node,
3790 Node* extra_slow_test,
3791 Node* *return_size_val,
3792 bool deoptimize_on_exception) {
3793 // Compute size in doublewords
3794 // The size is always an integral number of doublewords, represented
3795 // as a positive bytewise size stored in the klass's layout_helper.
3796 // The layout_helper also encodes (in a low bit) the need for a slow path.
3797 jint layout_con = Klass::_lh_neutral_value;
3798 Node* layout_val = get_layout_helper(klass_node, layout_con);
3799 int layout_is_con = (layout_val == nullptr);
3800
3801 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3802 // Generate the initial go-slow test. It's either ALWAYS (return a
3803 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3804 // case) a computed value derived from the layout_helper.
3805 Node* initial_slow_test = nullptr;
3806 if (layout_is_con) {
3807 assert(!StressReflectiveCode, "stress mode does not use these paths");
3808 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3809 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3810 } else { // reflective case
3811 // This reflective path is used by Unsafe.allocateInstance.
3812 // (It may be stress-tested by specifying StressReflectiveCode.)
3813 // Basically, we want to get into the VM is there's an illegal argument.
3814 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3815 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3816 if (extra_slow_test != intcon(0)) {
3817 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3818 }
3819 // (Macro-expander will further convert this to a Bool, if necessary.)
3830
3831 // Clear the low bits to extract layout_helper_size_in_bytes:
3832 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3833 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3834 size = _gvn.transform( new AndXNode(size, mask) );
3835 }
3836 if (return_size_val != nullptr) {
3837 (*return_size_val) = size;
3838 }
3839
3840 // This is a precise notnull oop of the klass.
3841 // (Actually, it need not be precise if this is a reflective allocation.)
3842 // It's what we cast the result to.
3843 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3844 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3845 const TypeOopPtr* oop_type = tklass->as_instance_type();
3846
3847 // Now generate allocation code
3848
3849 // The entire memory state is needed for slow path of the allocation
3850 // since GC and deoptimization can happened.
3851 Node *mem = reset_memory();
3852 set_all_memory(mem); // Create new memory state
3853
3854 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3855 control(), mem, i_o(),
3856 size, klass_node,
3857 initial_slow_test);
3858
3859 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3860 }
3861
3862 //-------------------------------new_array-------------------------------------
3863 // helper for both newarray and anewarray
3864 // The 'length' parameter is (obviously) the length of the array.
3865 // The optional arguments are for specialized use by intrinsics:
3866 // - If 'return_size_val', report the non-padded array size (sum of header size
3867 // and array body) to the caller.
3868 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3869 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3870 Node* length, // number of array elements
3871 int nargs, // number of arguments to push back for uncommon trap
3872 Node* *return_size_val,
3873 bool deoptimize_on_exception) {
3874 jint layout_con = Klass::_lh_neutral_value;
3875 Node* layout_val = get_layout_helper(klass_node, layout_con);
3876 int layout_is_con = (layout_val == nullptr);
3877
3878 if (!layout_is_con && !StressReflectiveCode &&
3879 !too_many_traps(Deoptimization::Reason_class_check)) {
3880 // This is a reflective array creation site.
3881 // Optimistically assume that it is a subtype of Object[],
3882 // so that we can fold up all the address arithmetic.
3883 layout_con = Klass::array_layout_helper(T_OBJECT);
3884 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3885 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3886 { BuildCutout unless(this, bol_lh, PROB_MAX);
3887 inc_sp(nargs);
3888 uncommon_trap(Deoptimization::Reason_class_check,
3889 Deoptimization::Action_maybe_recompile);
3890 }
3891 layout_val = nullptr;
3892 layout_is_con = true;
3893 }
3894
3895 // Generate the initial go-slow test. Make sure we do not overflow
3896 // if length is huge (near 2Gig) or negative! We do not need
3897 // exact double-words here, just a close approximation of needed
3898 // double-words. We can't add any offset or rounding bits, lest we
3899 // take a size -1 of bytes and make it positive. Use an unsigned
3900 // compare, so negative sizes look hugely positive.
3901 int fast_size_limit = FastAllocateSizeLimit;
3902 if (layout_is_con) {
3903 assert(!StressReflectiveCode, "stress mode does not use these paths");
3904 // Increase the size limit if we have exact knowledge of array type.
3905 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3906 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3907 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3908 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3909 }
3910
3911 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3912 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3913
3914 // --- Size Computation ---
3915 // array_size = round_to_heap(array_header + (length << elem_shift));
3916 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3917 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3918 // The rounding mask is strength-reduced, if possible.
3919 int round_mask = MinObjAlignmentInBytes - 1;
3920 Node* header_size = nullptr;
3921 // (T_BYTE has the weakest alignment and size restrictions...)
3922 if (layout_is_con) {
3923 int hsize = Klass::layout_helper_header_size(layout_con);
3924 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3925 if ((round_mask & ~right_n_bits(eshift)) == 0)
3926 round_mask = 0; // strength-reduce it if it goes away completely
3927 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3928 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3929 assert(header_size_min <= hsize, "generic minimum is smallest");
3930 header_size = intcon(hsize);
3931 } else {
3932 Node* hss = intcon(Klass::_lh_header_size_shift);
3933 Node* hsm = intcon(Klass::_lh_header_size_mask);
3934 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3935 header_size = _gvn.transform(new AndINode(header_size, hsm));
3936 }
3937
3938 Node* elem_shift = nullptr;
3939 if (layout_is_con) {
3940 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3941 if (eshift != 0)
3942 elem_shift = intcon(eshift);
3943 } else {
3944 // There is no need to mask or shift this value.
3945 // The semantics of LShiftINode include an implicit mask to 0x1F.
3946 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3947 elem_shift = layout_val;
3996 }
3997 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
3998
3999 if (return_size_val != nullptr) {
4000 // This is the size
4001 (*return_size_val) = non_rounded_size;
4002 }
4003
4004 Node* size = non_rounded_size;
4005 if (round_mask != 0) {
4006 Node* mask1 = MakeConX(round_mask);
4007 size = _gvn.transform(new AddXNode(size, mask1));
4008 Node* mask2 = MakeConX(~round_mask);
4009 size = _gvn.transform(new AndXNode(size, mask2));
4010 }
4011 // else if round_mask == 0, the size computation is self-rounding
4012
4013 // Now generate allocation code
4014
4015 // The entire memory state is needed for slow path of the allocation
4016 // since GC and deoptimization can happened.
4017 Node *mem = reset_memory();
4018 set_all_memory(mem); // Create new memory state
4019
4020 if (initial_slow_test->is_Bool()) {
4021 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4022 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4023 }
4024
4025 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
4026 Node* valid_length_test = _gvn.intcon(1);
4027 if (ary_type->isa_aryptr()) {
4028 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4029 jint max = TypeAryPtr::max_array_length(bt);
4030 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4031 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4032 }
4033
4034 // Create the AllocateArrayNode and its result projections
4035 AllocateArrayNode* alloc
4036 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4037 control(), mem, i_o(),
4038 size, klass_node,
4039 initial_slow_test,
4040 length, valid_length_test);
4041
4042 // Cast to correct type. Note that the klass_node may be constant or not,
4043 // and in the latter case the actual array type will be inexact also.
4044 // (This happens via a non-constant argument to inline_native_newArray.)
4045 // In any case, the value of klass_node provides the desired array type.
4046 const TypeInt* length_type = _gvn.find_int_type(length);
4047 if (ary_type->isa_aryptr() && length_type != nullptr) {
4048 // Try to get a better type than POS for the size
4049 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4050 }
4051
4052 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4053
4054 array_ideal_length(alloc, ary_type, true);
4055 return javaoop;
4056 }
4057
4058 // The following "Ideal_foo" functions are placed here because they recognize
4059 // the graph shapes created by the functions immediately above.
4060
4061 //---------------------------Ideal_allocation----------------------------------
4156 void GraphKit::add_parse_predicates(int nargs) {
4157 if (ShortRunningLongLoop) {
4158 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4159 // walking up from the loop.
4160 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4161 }
4162 if (UseLoopPredicate) {
4163 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4164 if (UseProfiledLoopPredicate) {
4165 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4166 }
4167 }
4168 if (UseAutoVectorizationPredicate) {
4169 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4170 }
4171 // Loop Limit Check Predicate should be near the loop.
4172 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4173 }
4174
4175 void GraphKit::sync_kit(IdealKit& ideal) {
4176 set_all_memory(ideal.merged_memory());
4177 set_i_o(ideal.i_o());
4178 set_control(ideal.ctrl());
4179 }
4180
4181 void GraphKit::final_sync(IdealKit& ideal) {
4182 // Final sync IdealKit and graphKit.
4183 sync_kit(ideal);
4184 }
4185
4186 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4187 Node* len = load_array_length(load_String_value(str, set_ctrl));
4188 Node* coder = load_String_coder(str, set_ctrl);
4189 // Divide length by 2 if coder is UTF16
4190 return _gvn.transform(new RShiftINode(len, coder));
4191 }
4192
4193 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4194 int value_offset = java_lang_String::value_offset();
4195 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4196 false, nullptr, 0);
4197 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4198 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4199 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4200 ciTypeArrayKlass::make(T_BYTE), true, 0);
4201 Node* p = basic_plus_adr(str, str, value_offset);
4202 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4203 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4204 return load;
4205 }
4206
4207 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4208 if (!CompactStrings) {
4209 return intcon(java_lang_String::CODER_UTF16);
4210 }
4211 int coder_offset = java_lang_String::coder_offset();
4212 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4213 false, nullptr, 0);
4214 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4215
4216 Node* p = basic_plus_adr(str, str, coder_offset);
4217 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4218 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4219 return load;
4220 }
4221
4222 void GraphKit::store_String_value(Node* str, Node* value) {
4223 int value_offset = java_lang_String::value_offset();
4224 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4225 false, nullptr, 0);
4226 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4227
4228 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4229 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4230 }
4231
4232 void GraphKit::store_String_coder(Node* str, Node* value) {
4233 int coder_offset = java_lang_String::coder_offset();
4234 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4235 false, nullptr, 0);
4236 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4237
4238 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4239 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4240 }
4241
4242 // Capture src and dst memory state with a MergeMemNode
4243 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4244 if (src_type == dst_type) {
4245 // Types are equal, we don't need a MergeMemNode
4246 return memory(src_type);
4247 }
4248 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4249 record_for_igvn(merge); // fold it up later, if possible
4250 int src_idx = C->get_alias_index(src_type);
4251 int dst_idx = C->get_alias_index(dst_type);
4252 merge->set_memory_at(src_idx, memory(src_idx));
4253 merge->set_memory_at(dst_idx, memory(dst_idx));
4254 return merge;
4255 }
4328 i_char->init_req(2, AddI(i_char, intcon(2)));
4329
4330 set_control(IfFalse(iff));
4331 set_memory(st, TypeAryPtr::BYTES);
4332 }
4333
4334 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4335 if (!field->is_constant()) {
4336 return nullptr; // Field not marked as constant.
4337 }
4338 ciInstance* holder = nullptr;
4339 if (!field->is_static()) {
4340 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4341 if (const_oop != nullptr && const_oop->is_instance()) {
4342 holder = const_oop->as_instance();
4343 }
4344 }
4345 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4346 /*is_unsigned_load=*/false);
4347 if (con_type != nullptr) {
4348 return makecon(con_type);
4349 }
4350 return nullptr;
4351 }
4352
4353 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4354 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4355 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4356 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4357 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4358 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4359 return casted_obj;
4360 }
4361 return obj;
4362 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/parse.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/runtime.hpp"
55 #include "opto/subtypenode.hpp"
56 #include "runtime/arguments.hpp"
57 #include "runtime/deoptimization.hpp"
58 #include "runtime/sharedRuntime.hpp"
59 #include "runtime/stubRoutines.hpp"
60 #include "utilities/bitMap.inline.hpp"
61 #include "utilities/growableArray.hpp"
62 #include "utilities/powerOfTwo.hpp"
63
64 //----------------------------GraphKit-----------------------------------------
65 // Main utility constructor.
66 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
67 : Phase(Phase::Parser),
68 _env(C->env()),
69 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
70 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
71 {
72 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
73 _exceptions = jvms->map()->next_exception();
74 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
75 set_jvms(jvms);
76 #ifdef ASSERT
77 if (_gvn.is_IterGVN() != nullptr) {
78 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
79 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
80 _worklist_size = _gvn.C->igvn_worklist()->size();
81 }
82 #endif
83 }
84
85 // Private constructor for parser.
86 GraphKit::GraphKit()
87 : Phase(Phase::Parser),
88 _env(C->env()),
89 _gvn(*C->initial_gvn()),
90 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
91 {
92 _exceptions = nullptr;
93 set_map(nullptr);
94 DEBUG_ONLY(_sp = -99);
95 DEBUG_ONLY(set_bci(-99));
96 }
97
98 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
99 : Phase(Phase::Parser),
100 _env(C->env()),
101 _gvn(igvn),
102 _exceptions(nullptr),
103 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
104 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
105 assert(sft->next_exception() == nullptr, "must not have a pending exception");
106 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
107 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
108 for (uint i = 0; i < sft->req(); i++) {
109 cloned_map->init_req(i, sft->in(i));
110 }
111 igvn.record_for_igvn(cloned_map);
112 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
113 current->set_map(cloned_map);
114 }
115 set_jvms(cloned_jvms);
116 set_all_memory(reset_memory());
117 }
118
119 //---------------------------clean_stack---------------------------------------
120 // Clear away rubbish from the stack area of the JVM state.
121 // This destroys any arguments that may be waiting on the stack.
122 void GraphKit::clean_stack(int from_sp) {
123 SafePointNode* map = this->map();
124 JVMState* jvms = this->jvms();
125 int stk_size = jvms->stk_size();
126 int stkoff = jvms->stkoff();
127 Node* top = this->top();
128 for (int i = from_sp; i < stk_size; i++) {
129 if (map->in(stkoff + i) != top) {
130 map->set_req(stkoff + i, top);
131 }
132 }
133 }
134
135
136 //--------------------------------sync_jvms-----------------------------------
137 // Make sure our current jvms agrees with our parse state.
366 }
367 static inline void add_one_req(Node* dstphi, Node* src) {
368 assert(is_hidden_merge(dstphi), "must be a special merge node");
369 assert(!is_hidden_merge(src), "must not be a special merge node");
370 dstphi->add_req(src);
371 }
372
373 //-----------------------combine_exception_states------------------------------
374 // This helper function combines exception states by building phis on a
375 // specially marked state-merging region. These regions and phis are
376 // untransformed, and can build up gradually. The region is marked by
377 // having a control input of its exception map, rather than null. Such
378 // regions do not appear except in this function, and in use_exception_state.
379 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
380 if (failing_internal()) {
381 return; // dying anyway...
382 }
383 JVMState* ex_jvms = ex_map->_jvms;
384 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
385 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
386 // TODO 8325632 Re-enable
387 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
388 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
389 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
390 assert(ex_map->req() == phi_map->req(), "matching maps");
391 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
392 Node* hidden_merge_mark = root();
393 Node* region = phi_map->control();
394 MergeMemNode* phi_mem = phi_map->merged_memory();
395 MergeMemNode* ex_mem = ex_map->merged_memory();
396 if (region->in(0) != hidden_merge_mark) {
397 // The control input is not (yet) a specially-marked region in phi_map.
398 // Make it so, and build some phis.
399 region = new RegionNode(2);
400 _gvn.set_type(region, Type::CONTROL);
401 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
402 region->init_req(1, phi_map->control());
403 phi_map->set_control(region);
404 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
405 record_for_igvn(io_phi);
406 _gvn.set_type(io_phi, Type::ABIO);
407 phi_map->set_i_o(io_phi);
937 if (PrintMiscellaneous && (Verbose || WizardMode)) {
938 tty->print_cr("Zombie local %d: ", local);
939 jvms->dump();
940 }
941 return false;
942 }
943 }
944 }
945 return true;
946 }
947
948 #endif //ASSERT
949
950 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
951 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
952 ciMethod* cur_method = jvms->method();
953 int cur_bci = jvms->bci();
954 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
955 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
956 return Interpreter::bytecode_should_reexecute(code) ||
957 (is_anewarray && (code == Bytecodes::_multianewarray));
958 // Reexecute _multianewarray bytecode which was replaced with
959 // sequence of [a]newarray. See Parse::do_multianewarray().
960 //
961 // Note: interpreter should not have it set since this optimization
962 // is limited by dimensions and guarded by flag so in some cases
963 // multianewarray() runtime calls will be generated and
964 // the bytecode should not be reexecutes (stack will not be reset).
965 } else {
966 return false;
967 }
968 }
969
970 // Helper function for adding JVMState and debug information to node
971 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
972 // Add the safepoint edges to the call (or other safepoint).
973
974 // Make sure dead locals are set to top. This
975 // should help register allocation time and cut down on the size
976 // of the deoptimization information.
977 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1005
1006 if (env()->should_retain_local_variables()) {
1007 // At any safepoint, this method can get breakpointed, which would
1008 // then require an immediate deoptimization.
1009 can_prune_locals = false; // do not prune locals
1010 stack_slots_not_pruned = 0;
1011 }
1012
1013 // do not scribble on the input jvms
1014 JVMState* out_jvms = youngest_jvms->clone_deep(C);
1015 call->set_jvms(out_jvms); // Start jvms list for call node
1016
1017 // For a known set of bytecodes, the interpreter should reexecute them if
1018 // deoptimization happens. We set the reexecute state for them here
1019 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
1020 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
1021 #ifdef ASSERT
1022 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
1023 assert(method() == youngest_jvms->method(), "sanity");
1024 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
1025 // TODO 8371125
1026 // assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
1027 #endif // ASSERT
1028 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
1029 }
1030
1031 // Presize the call:
1032 DEBUG_ONLY(uint non_debug_edges = call->req());
1033 call->add_req_batch(top(), youngest_jvms->debug_depth());
1034 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
1035
1036 // Set up edges so that the call looks like this:
1037 // Call [state:] ctl io mem fptr retadr
1038 // [parms:] parm0 ... parmN
1039 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1040 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
1041 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
1042 // Note that caller debug info precedes callee debug info.
1043
1044 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
1045 uint debug_ptr = call->req();
1046
1047 // Loop over the map input edges associated with jvms, add them
1048 // to the call node, & reset all offsets to match call node array.
1049
1050 JVMState* callee_jvms = nullptr;
1051 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
1052 uint debug_end = debug_ptr;
1053 uint debug_start = debug_ptr - in_jvms->debug_size();
1054 debug_ptr = debug_start; // back up the ptr
1055
1056 uint p = debug_start; // walks forward in [debug_start, debug_end)
1057 uint j, k, l;
1058 SafePointNode* in_map = in_jvms->map();
1059 out_jvms->set_map(call);
1060
1061 if (can_prune_locals) {
1062 assert(in_jvms->method() == out_jvms->method(), "sanity");
1063 // If the current throw can reach an exception handler in this JVMS,
1064 // then we must keep everything live that can reach that handler.
1065 // As a quick and dirty approximation, we look for any handlers at all.
1066 if (in_jvms->method()->has_exception_handlers()) {
1067 can_prune_locals = false;
1068 }
1069 }
1070
1071 // Add the Locals
1072 k = in_jvms->locoff();
1073 l = in_jvms->loc_size();
1074 out_jvms->set_locoff(p);
1075 if (!can_prune_locals) {
1076 for (j = 0; j < l; j++) {
1077 call->set_req(p++, in_map->in(k + j));
1078 }
1079 } else {
1080 p += l; // already set to top above by add_req_batch
1081 }
1082
1083 // Add the Expression Stack
1084 k = in_jvms->stkoff();
1085 l = in_jvms->sp();
1086 out_jvms->set_stkoff(p);
1087 if (!can_prune_locals) {
1088 for (j = 0; j < l; j++) {
1089 call->set_req(p++, in_map->in(k + j));
1090 }
1091 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1092 // Divide stack into {S0,...,S1}, where S0 is set to top.
1093 uint s1 = stack_slots_not_pruned;
1094 stack_slots_not_pruned = 0; // for next iteration
1095 if (s1 > l) s1 = l;
1096 uint s0 = l - s1;
1097 p += s0; // skip the tops preinstalled by add_req_batch
1098 for (j = s0; j < l; j++)
1099 call->set_req(p++, in_map->in(k+j));
1100 } else {
1101 p += l; // already set to top above by add_req_batch
1102 }
1103
1104 // Add the Monitors
1105 k = in_jvms->monoff();
1106 l = in_jvms->mon_size();
1107 out_jvms->set_monoff(p);
1108 for (j = 0; j < l; j++)
1109 call->set_req(p++, in_map->in(k+j));
1110
1111 // Copy any scalar object fields.
1112 k = in_jvms->scloff();
1113 l = in_jvms->scl_size();
1114 out_jvms->set_scloff(p);
1115 for (j = 0; j < l; j++)
1116 call->set_req(p++, in_map->in(k+j));
1117
1118 // Finish the new jvms.
1119 out_jvms->set_endoff(p);
1120
1121 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1122 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1123 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1124 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1125 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1126 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1127
1128 // Update the two tail pointers in parallel.
1129 callee_jvms = out_jvms;
1130 out_jvms = out_jvms->caller();
1131 in_jvms = in_jvms->caller();
1132 }
1133
1134 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1135
1136 // Test the correctness of JVMState::debug_xxx accessors:
1137 assert(call->jvms()->debug_start() == non_debug_edges, "");
1138 assert(call->jvms()->debug_end() == call->req(), "");
1139 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1140 }
1141
1142 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1143 Bytecodes::Code code = java_bc();
1144 if (code == Bytecodes::_wide) {
1145 code = method()->java_code_at_bci(bci() + 1);
1146 }
1147
1148 if (code != Bytecodes::_illegal) {
1149 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1299 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1300 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1301 return _gvn.transform( new AndLNode(conv, mask) );
1302 }
1303
1304 Node* GraphKit::ConvL2I(Node* offset) {
1305 // short-circuit a common case
1306 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1307 if (offset_con != (jlong)Type::OffsetBot) {
1308 return intcon((int) offset_con);
1309 }
1310 return _gvn.transform( new ConvL2INode(offset));
1311 }
1312
1313 //-------------------------load_object_klass-----------------------------------
1314 Node* GraphKit::load_object_klass(Node* obj) {
1315 // Special-case a fresh allocation to avoid building nodes:
1316 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1317 if (akls != nullptr) return akls;
1318 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1319 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1320 }
1321
1322 //-------------------------load_array_length-----------------------------------
1323 Node* GraphKit::load_array_length(Node* array) {
1324 // Special-case a fresh allocation to avoid building nodes:
1325 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1326 Node *alen;
1327 if (alloc == nullptr) {
1328 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1329 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1330 } else {
1331 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1332 }
1333 return alen;
1334 }
1335
1336 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1337 const TypeOopPtr* oop_type,
1338 bool replace_length_in_map) {
1339 Node* length = alloc->Ideal_length();
1348 replace_in_map(length, ccast);
1349 }
1350 return ccast;
1351 }
1352 }
1353 return length;
1354 }
1355
1356 //------------------------------do_null_check----------------------------------
1357 // Helper function to do a null pointer check. Returned value is
1358 // the incoming address with null casted away. You are allowed to use the
1359 // not-null value only if you are control dependent on the test.
1360 #ifndef PRODUCT
1361 extern uint explicit_null_checks_inserted,
1362 explicit_null_checks_elided;
1363 #endif
1364 Node* GraphKit::null_check_common(Node* value, BasicType type,
1365 // optional arguments for variations:
1366 bool assert_null,
1367 Node* *null_control,
1368 bool speculative,
1369 bool null_marker_check) {
1370 assert(!assert_null || null_control == nullptr, "not both at once");
1371 if (stopped()) return top();
1372 NOT_PRODUCT(explicit_null_checks_inserted++);
1373
1374 if (value->is_InlineType()) {
1375 // Null checking a scalarized but nullable inline type. Check the null marker
1376 // input instead of the oop input to avoid keeping buffer allocations alive.
1377 InlineTypeNode* vtptr = value->as_InlineType();
1378 while (vtptr->get_oop()->is_InlineType()) {
1379 vtptr = vtptr->get_oop()->as_InlineType();
1380 }
1381 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1382 if (stopped()) {
1383 return top();
1384 }
1385 if (assert_null) {
1386 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1387 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1388 // replace_in_map(value, vtptr);
1389 // return vtptr;
1390 replace_in_map(value, null());
1391 return null();
1392 }
1393 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1394 return cast_not_null(value, do_replace_in_map);
1395 }
1396
1397 // Construct null check
1398 Node *chk = nullptr;
1399 switch(type) {
1400 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1401 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1402 case T_ARRAY : // fall through
1403 type = T_OBJECT; // simplify further tests
1404 case T_OBJECT : {
1405 const Type *t = _gvn.type( value );
1406
1407 const TypeOopPtr* tp = t->isa_oopptr();
1408 if (tp != nullptr && !tp->is_loaded()
1409 // Only for do_null_check, not any of its siblings:
1410 && !assert_null && null_control == nullptr) {
1411 // Usually, any field access or invocation on an unloaded oop type
1412 // will simply fail to link, since the statically linked class is
1413 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1414 // the static class is loaded but the sharper oop type is not.
1415 // Rather than checking for this obscure case in lots of places,
1416 // we simply observe that a null check on an unloaded class
1480 }
1481 Node *oldcontrol = control();
1482 set_control(cfg);
1483 Node *res = cast_not_null(value);
1484 set_control(oldcontrol);
1485 NOT_PRODUCT(explicit_null_checks_elided++);
1486 return res;
1487 }
1488 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1489 if (cfg == nullptr) break; // Quit at region nodes
1490 depth++;
1491 }
1492 }
1493
1494 //-----------
1495 // Branch to failure if null
1496 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1497 Deoptimization::DeoptReason reason;
1498 if (assert_null) {
1499 reason = Deoptimization::reason_null_assert(speculative);
1500 } else if (type == T_OBJECT || null_marker_check) {
1501 reason = Deoptimization::reason_null_check(speculative);
1502 } else {
1503 reason = Deoptimization::Reason_div0_check;
1504 }
1505 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1506 // ciMethodData::has_trap_at will return a conservative -1 if any
1507 // must-be-null assertion has failed. This could cause performance
1508 // problems for a method after its first do_null_assert failure.
1509 // Consider using 'Reason_class_check' instead?
1510
1511 // To cause an implicit null check, we set the not-null probability
1512 // to the maximum (PROB_MAX). For an explicit check the probability
1513 // is set to a smaller value.
1514 if (null_control != nullptr || too_many_traps(reason)) {
1515 // probability is less likely
1516 ok_prob = PROB_LIKELY_MAG(3);
1517 } else if (!assert_null &&
1518 (ImplicitNullCheckThreshold > 0) &&
1519 method() != nullptr &&
1520 (method()->method_data()->trap_count(reason)
1554 }
1555
1556 if (assert_null) {
1557 // Cast obj to null on this path.
1558 replace_in_map(value, zerocon(type));
1559 return zerocon(type);
1560 }
1561
1562 // Cast obj to not-null on this path, if there is no null_control.
1563 // (If there is a null_control, a non-null value may come back to haunt us.)
1564 if (type == T_OBJECT) {
1565 Node* cast = cast_not_null(value, false);
1566 if (null_control == nullptr || (*null_control) == top())
1567 replace_in_map(value, cast);
1568 value = cast;
1569 }
1570
1571 return value;
1572 }
1573
1574 //------------------------------cast_not_null----------------------------------
1575 // Cast obj to not-null on this path
1576 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1577 if (obj->is_InlineType()) {
1578 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1579 vt->as_InlineType()->set_null_marker(_gvn);
1580 vt = _gvn.transform(vt);
1581 if (do_replace_in_map) {
1582 replace_in_map(obj, vt);
1583 }
1584 return vt;
1585 }
1586 const Type *t = _gvn.type(obj);
1587 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1588 // Object is already not-null?
1589 if( t == t_not_null ) return obj;
1590
1591 Node* cast = new CastPPNode(control(), obj,t_not_null);
1592 cast = _gvn.transform( cast );
1593
1594 // Scan for instances of 'obj' in the current JVM mapping.
1595 // These instances are known to be not-null after the test.
1596 if (do_replace_in_map)
1597 replace_in_map(obj, cast);
1598
1599 return cast; // Return casted value
1600 }
1601
1602 // Sometimes in intrinsics, we implicitly know an object is not null
1603 // (there's no actual null check) so we can cast it to not null. In
1604 // the course of optimizations, the input to the cast can become null.
1605 // In that case that data path will die and we need the control path
1660 Node* GraphKit::memory(uint alias_idx) {
1661 MergeMemNode* mem = merged_memory();
1662 Node* p = mem->memory_at(alias_idx);
1663 assert(p != mem->empty_memory(), "empty");
1664 _gvn.set_type(p, Type::MEMORY); // must be mapped
1665 return p;
1666 }
1667
1668 //-----------------------------reset_memory------------------------------------
1669 Node* GraphKit::reset_memory() {
1670 Node* mem = map()->memory();
1671 // do not use this node for any more parsing!
1672 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1673 return _gvn.transform( mem );
1674 }
1675
1676 //------------------------------set_all_memory---------------------------------
1677 void GraphKit::set_all_memory(Node* newmem) {
1678 Node* mergemem = MergeMemNode::make(newmem);
1679 gvn().set_type_bottom(mergemem);
1680 if (_gvn.is_IterGVN() != nullptr) {
1681 record_for_igvn(mergemem);
1682 }
1683 map()->set_memory(mergemem);
1684 }
1685
1686 //------------------------------set_all_memory_call----------------------------
1687 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1688 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1689 set_all_memory(newmem);
1690 }
1691
1692 //=============================================================================
1693 //
1694 // parser factory methods for MemNodes
1695 //
1696 // These are layered on top of the factory methods in LoadNode and StoreNode,
1697 // and integrate with the parser's memory state and _gvn engine.
1698 //
1699
1700 // factory methods in "int adr_idx"
1701 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1702 MemNode::MemOrd mo,
1703 LoadNode::ControlDependency control_dependency,
1704 bool require_atomic_access,
1705 bool unaligned,
1706 bool mismatched,
1707 bool unsafe,
1708 uint8_t barrier_data) {
1709 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1710 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1711 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1712 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1713 Node* mem = memory(adr_idx);
1714 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1715 ld = _gvn.transform(ld);
1716
1717 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1718 // Improve graph before escape analysis and boxing elimination.
1719 record_for_igvn(ld);
1720 if (ld->is_DecodeN()) {
1721 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1722 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1723 // a Phi). Recording such cases is still perfectly sound, but may be
1724 // unnecessary and result in some minor IGVN overhead.
1725 record_for_igvn(ld->in(1));
1726 }
1727 }
1728 return ld;
1729 }
1730
1731 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1732 MemNode::MemOrd mo,
1733 bool require_atomic_access,
1734 bool unaligned,
1735 bool mismatched,
1736 bool unsafe,
1750 if (unsafe) {
1751 st->as_Store()->set_unsafe_access();
1752 }
1753 st->as_Store()->set_barrier_data(barrier_data);
1754 st = _gvn.transform(st);
1755 set_memory(st, adr_idx);
1756 // Back-to-back stores can only remove intermediate store with DU info
1757 // so push on worklist for optimizer.
1758 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1759 record_for_igvn(st);
1760
1761 return st;
1762 }
1763
1764 Node* GraphKit::access_store_at(Node* obj,
1765 Node* adr,
1766 const TypePtr* adr_type,
1767 Node* val,
1768 const Type* val_type,
1769 BasicType bt,
1770 DecoratorSet decorators,
1771 bool safe_for_replace,
1772 const InlineTypeNode* vt) {
1773 // Transformation of a value which could be null pointer (CastPP #null)
1774 // could be delayed during Parse (for example, in adjust_map_after_if()).
1775 // Execute transformation here to avoid barrier generation in such case.
1776 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1777 val = _gvn.makecon(TypePtr::NULL_PTR);
1778 }
1779
1780 if (stopped()) {
1781 return top(); // Dead path ?
1782 }
1783
1784 assert(val != nullptr, "not dead path");
1785 if (val->is_InlineType()) {
1786 // Store to non-flat field. Buffer the inline type and make sure
1787 // the store is re-executed if the allocation triggers deoptimization.
1788 PreserveReexecuteState preexecs(this);
1789 jvms()->set_should_reexecute(true);
1790 val = val->as_InlineType()->buffer(this, safe_for_replace);
1791 }
1792
1793 C2AccessValuePtr addr(adr, adr_type);
1794 C2AccessValue value(val, val_type);
1795 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1796 if (access.is_raw()) {
1797 return _barrier_set->BarrierSetC2::store_at(access, value);
1798 } else {
1799 return _barrier_set->store_at(access, value);
1800 }
1801 }
1802
1803 Node* GraphKit::access_load_at(Node* obj, // containing obj
1804 Node* adr, // actual address to store val at
1805 const TypePtr* adr_type,
1806 const Type* val_type,
1807 BasicType bt,
1808 DecoratorSet decorators,
1809 Node* ctl) {
1810 if (stopped()) {
1811 return top(); // Dead path ?
1812 }
1813
1814 SavedState old_state(this);
1815 C2AccessValuePtr addr(adr, adr_type);
1816 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1817 Node* load;
1818 if (access.is_raw()) {
1819 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1820 } else {
1821 load = _barrier_set->load_at(access, val_type);
1822 }
1823
1824 // Restore the previous state only if the load got folded to a constant
1825 // and we can discard any barriers that might have been added.
1826 if (load == nullptr || !load->is_Con()) {
1827 old_state.discard();
1828 }
1829 return load;
1830 }
1831
1832 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1833 const Type* val_type,
1834 BasicType bt,
1835 DecoratorSet decorators) {
1836 if (stopped()) {
1918 Node* new_val,
1919 const Type* value_type,
1920 BasicType bt,
1921 DecoratorSet decorators) {
1922 C2AccessValuePtr addr(adr, adr_type);
1923 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1924 if (access.is_raw()) {
1925 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1926 } else {
1927 return _barrier_set->atomic_add_at(access, new_val, value_type);
1928 }
1929 }
1930
1931 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1932 return _barrier_set->clone(this, src, dst, size, is_array);
1933 }
1934
1935 //-------------------------array_element_address-------------------------
1936 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1937 const TypeInt* sizetype, Node* ctrl) {
1938 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1939 uint shift;
1940 uint header;
1941 if (arytype->is_flat() && arytype->klass_is_exact()) {
1942 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1943 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1944 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1945 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1946 // though we don't need the address node in this case and throw it away again.
1947 shift = arytype->flat_log_elem_size();
1948 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1949 } else {
1950 shift = exact_log2(type2aelembytes(elembt));
1951 header = arrayOopDesc::base_offset_in_bytes(elembt);
1952 }
1953
1954 // short-circuit a common case (saves lots of confusing waste motion)
1955 jint idx_con = find_int_con(idx, -1);
1956 if (idx_con >= 0) {
1957 intptr_t offset = header + ((intptr_t)idx_con << shift);
1958 return basic_plus_adr(ary, offset);
1959 }
1960
1961 // must be correct type for alignment purposes
1962 Node* base = basic_plus_adr(ary, header);
1963 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1964 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1965 return basic_plus_adr(ary, base, scale);
1966 }
1967
1968 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1969 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1970 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1971 return cast_to_flat_array_exact(array, elem_vk, true, false);
1972 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1973 return cast_to_flat_array_exact(array, elem_vk, true, true);
1974 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1975 return cast_to_flat_array_exact(array, elem_vk, false, true);
1976 }
1977
1978 bool is_null_free = false;
1979 if (!elem_vk->has_nullable_atomic_layout()) {
1980 // Element does not have a nullable flat layout, cannot be nullable
1981 is_null_free = true;
1982 }
1983
1984 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1985 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1986 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1987 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1988 }
1989
1990 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1991 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1992 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1993 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1994 assert(arytype->klass_is_exact(), "inconsistency");
1995 assert(arytype->is_flat(), "inconsistency");
1996 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1997 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1998 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1999 }
2000
2001 //-------------------------load_array_element-------------------------
2002 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
2003 const Type* elemtype = arytype->elem();
2004 BasicType elembt = elemtype->array_element_basic_type();
2005 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
2006 if (elembt == T_NARROWOOP) {
2007 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
2008 }
2009 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2010 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2011 return ld;
2012 }
2013
2014 //-------------------------set_arguments_for_java_call-------------------------
2015 // Arguments (pre-popped from the stack) are taken from the JVMS.
2016 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2017 PreserveReexecuteState preexecs(this);
2018 if (Arguments::is_valhalla_enabled()) {
2019 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2020 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2021 jvms()->set_should_reexecute(true);
2022 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2023 inc_sp(arg_size);
2024 }
2025 // Add the call arguments
2026 const TypeTuple* domain = call->tf()->domain_sig();
2027 uint nargs = domain->cnt();
2028 int arg_num = 0;
2029 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2030 uint arg_idx = i - TypeFunc::Parms;
2031 Node* arg = argument(arg_idx);
2032 const Type* t = domain->field_at(i);
2033 // TODO 8284443 A static call to a mismatched method should still be scalarized
2034 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2035 // We don't pass inline type arguments by reference but instead pass each field of the inline type
2036 if (!arg->is_InlineType()) {
2037 // There are 2 cases in which the argument has not been scalarized
2038 if (_gvn.type(arg)->is_zero_type()) {
2039 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2040 } else {
2041 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2042 // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2043 // type must be an exact value class. The method is devirtualized, and replaced with
2044 // a direct call with a scalarized receiver instead.
2045 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2046 assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2047 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2048 }
2049 }
2050 InlineTypeNode* vt = arg->as_InlineType();
2051 vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2052 // If an inline type argument is passed as fields, attach the Method* to the call site
2053 // to be able to access the extended signature later via attached_method_before_pc().
2054 // For example, see CompiledMethod::preserve_callee_argument_oops().
2055 call->set_override_symbolic_info(true);
2056 // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2057 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2058 C->dependencies()->assert_mismatch_calling_convention(call->method());
2059 arg_num++;
2060 continue;
2061 } else if (arg->is_InlineType()) {
2062 // Pass inline type argument via oop to callee
2063 arg = arg->as_InlineType()->buffer(this, true);
2064 }
2065 if (t != Type::HALF) {
2066 arg_num++;
2067 }
2068 call->init_req(idx++, arg);
2069 }
2070 }
2071
2072 //---------------------------set_edges_for_java_call---------------------------
2073 // Connect a newly created call into the current JVMS.
2074 // A return value node (if any) is returned from set_edges_for_java_call.
2075 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2076
2077 // Add the predefined inputs:
2078 call->init_req( TypeFunc::Control, control() );
2079 call->init_req( TypeFunc::I_O , i_o() );
2080 call->init_req( TypeFunc::Memory , reset_memory() );
2081 call->init_req( TypeFunc::FramePtr, frameptr() );
2082 call->init_req( TypeFunc::ReturnAdr, top() );
2083
2084 add_safepoint_edges(call, must_throw);
2085
2086 Node* xcall = _gvn.transform(call);
2087
2088 if (xcall == top()) {
2089 set_control(top());
2090 return;
2091 }
2092 assert(xcall == call, "call identity is stable");
2093
2094 // Re-use the current map to produce the result.
2095
2096 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2097 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2098 set_all_memory_call(xcall, separate_io_proj);
2099
2100 //return xcall; // no need, caller already has it
2101 }
2102
2103 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2104 if (stopped()) return top(); // maybe the call folded up?
2105
2106 // Note: Since any out-of-line call can produce an exception,
2107 // we always insert an I_O projection from the call into the result.
2108
2109 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2110
2111 if (separate_io_proj) {
2112 // The caller requested separate projections be used by the fall
2113 // through and exceptional paths, so replace the projections for
2114 // the fall through path.
2115 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2116 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2117 }
2118
2119 // Capture the return value, if any.
2120 Node* ret;
2121 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2122 ret = top();
2123 } else if (call->tf()->returns_inline_type_as_fields()) {
2124 // Return of multiple values (inline type fields): we create a
2125 // InlineType node, each field is a projection from the call.
2126 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2127 uint base_input = TypeFunc::Parms;
2128 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2129 } else {
2130 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2131 ciType* t = call->method()->return_type();
2132 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2133 // The return type is unloaded but the callee might later be C2 compiled and then return
2134 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2135 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2136 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2137 IdealKit ideal(this);
2138 IdealVariable res(ideal);
2139 ideal.declarations_done();
2140 // Change return type of call to scalarized return
2141 const TypeFunc* tf = call->_tf;
2142 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2143 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2144 call->_tf = new_tf;
2145 _gvn.set_type(call, call->Value(&_gvn));
2146 _gvn.set_type(ret, ret->Value(&_gvn));
2147 // Don't add store to buffer call if we are strength reducing
2148 if (!C->strength_reduction()) {
2149 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2150 // Return value is null
2151 ideal.set(res, makecon(TypePtr::NULL_PTR));
2152 } ideal.else_(); {
2153 // Return value is non-null
2154 sync_kit(ideal);
2155
2156 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2157 OptoRuntime::store_inline_type_fields_Type(),
2158 StubRoutines::store_inline_type_fields_to_buf(),
2159 nullptr, TypePtr::BOTTOM, ret);
2160
2161 // We don't know how many values are returned. This assumes the
2162 // worst case, that all available registers are used.
2163 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2164 if (domain->field_at(i) == Type::HALF) {
2165 store_to_buf_call->init_req(i, top());
2166 continue;
2167 }
2168 Node* proj =_gvn.transform(new ProjNode(call, i));
2169 store_to_buf_call->init_req(i, proj);
2170 }
2171 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2172
2173 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2174 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2175 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2176
2177 ideal.set(res, buf);
2178 ideal.sync_kit(this);
2179 } ideal.end_if();
2180 } else {
2181 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2182 Node* proj =_gvn.transform(new ProjNode(call, i));
2183 }
2184 ideal.set(res, ret);
2185 }
2186 sync_kit(ideal);
2187 ret = _gvn.transform(ideal.value(res));
2188 } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2189 // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2190 // result of merging several call paths. If one of them is made of an actual call node that
2191 // returns an oop, we need to call make_from_oop here as well because we want InlineType
2192 // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2193 ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2194 }
2195 }
2196
2197 return ret;
2198 }
2199
2200 //--------------------set_predefined_input_for_runtime_call--------------------
2201 // Reading and setting the memory state is way conservative here.
2202 // The real problem is that I am not doing real Type analysis on memory,
2203 // so I cannot distinguish card mark stores from other stores. Across a GC
2204 // point the Store Barrier and the card mark memory has to agree. I cannot
2205 // have a card mark store and its barrier split across the GC point from
2206 // either above or below. Here I get that to happen by reading ALL of memory.
2207 // A better answer would be to separate out card marks from other memory.
2208 // For now, return the input memory state, so that it can be reused
2209 // after the call, if this call has restricted memory effects.
2210 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2211 // Set fixed predefined input arguments
2212 call->init_req(TypeFunc::Control, control());
2213 call->init_req(TypeFunc::I_O, top()); // does no i/o
2214 call->init_req(TypeFunc::ReturnAdr, top());
2215 if (call->is_CallLeafPure()) {
2216 call->init_req(TypeFunc::Memory, top());
2278 if (use->is_MergeMem()) {
2279 wl.push(use);
2280 }
2281 }
2282 }
2283
2284 // Replace the call with the current state of the kit.
2285 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2286 JVMState* ejvms = nullptr;
2287 if (has_exceptions()) {
2288 ejvms = transfer_exceptions_into_jvms();
2289 }
2290
2291 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2292 ReplacedNodes replaced_nodes_exception;
2293 Node* ex_ctl = top();
2294
2295 SafePointNode* final_state = stop();
2296
2297 // Find all the needed outputs of this call
2298 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2299
2300 Unique_Node_List wl;
2301 Node* init_mem = call->in(TypeFunc::Memory);
2302 Node* final_mem = final_state->in(TypeFunc::Memory);
2303 Node* final_ctl = final_state->in(TypeFunc::Control);
2304 Node* final_io = final_state->in(TypeFunc::I_O);
2305
2306 // Replace all the old call edges with the edges from the inlining result
2307 if (callprojs->fallthrough_catchproj != nullptr) {
2308 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2309 }
2310 if (callprojs->fallthrough_memproj != nullptr) {
2311 if (final_mem->is_MergeMem()) {
2312 // Parser's exits MergeMem was not transformed but may be optimized
2313 final_mem = _gvn.transform(final_mem);
2314 }
2315 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2316 add_mergemem_users_to_worklist(wl, final_mem);
2317 }
2318 if (callprojs->fallthrough_ioproj != nullptr) {
2319 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2320 }
2321
2322 // Replace the result with the new result if it exists and is used
2323 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2324 // If the inlined code is dead, the result projections for an inline type returned as
2325 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2326 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2327 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2328 "unexpected number of results");
2329 // If we are doing strength reduction and the return type is not loaded we
2330 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2331 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2332 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2333 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2334 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2335 }
2336 } else {
2337 C->gvn_replace_by(callprojs->resproj[0], result);
2338 }
2339 }
2340
2341 if (ejvms == nullptr) {
2342 // No exception edges to simply kill off those paths
2343 if (callprojs->catchall_catchproj != nullptr) {
2344 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2345 }
2346 if (callprojs->catchall_memproj != nullptr) {
2347 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2348 }
2349 if (callprojs->catchall_ioproj != nullptr) {
2350 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2351 }
2352 // Replace the old exception object with top
2353 if (callprojs->exobj != nullptr) {
2354 C->gvn_replace_by(callprojs->exobj, C->top());
2355 }
2356 } else {
2357 GraphKit ekit(ejvms);
2358
2359 // Load my combined exception state into the kit, with all phis transformed:
2360 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2361 replaced_nodes_exception = ex_map->replaced_nodes();
2362
2363 Node* ex_oop = ekit.use_exception_state(ex_map);
2364
2365 if (callprojs->catchall_catchproj != nullptr) {
2366 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2367 ex_ctl = ekit.control();
2368 }
2369 if (callprojs->catchall_memproj != nullptr) {
2370 Node* ex_mem = ekit.reset_memory();
2371 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2372 add_mergemem_users_to_worklist(wl, ex_mem);
2373 }
2374 if (callprojs->catchall_ioproj != nullptr) {
2375 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2376 }
2377
2378 // Replace the old exception object with the newly created one
2379 if (callprojs->exobj != nullptr) {
2380 C->gvn_replace_by(callprojs->exobj, ex_oop);
2381 }
2382 }
2383
2384 // Disconnect the call from the graph
2385 call->disconnect_inputs(C);
2386 C->gvn_replace_by(call, C->top());
2387
2388 // Clean up any MergeMems that feed other MergeMems since the
2389 // optimizer doesn't like that.
2390 while (wl.size() > 0) {
2391 _gvn.transform(wl.pop());
2392 }
2393
2394 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2395 replaced_nodes.apply(C, final_ctl);
2396 }
2397 if (!ex_ctl->is_top() && do_replaced_nodes) {
2398 replaced_nodes_exception.apply(C, ex_ctl);
2399 }
2400 }
2401
2402
2403 //------------------------------increment_counter------------------------------
2404 // for statistics: increment a VM counter by 1
2405
2406 void GraphKit::increment_counter(address counter_addr) {
2407 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2408 increment_counter(adr1);
2409 }
2410
2411 void GraphKit::increment_counter(Node* counter_addr) {
2412 Node* ctrl = control();
2413 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2414 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2415 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2416 }
2417
2418 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2419 Node* halt = new HaltNode(ctrl, frameptr, reason
2420 PRODUCT_ONLY(COMMA generate_code_in_product));
2421 halt = _gvn.transform(halt);
2422 root()->add_req(halt);
2423 if (_gvn.is_IterGVN() != nullptr) {
2424 record_for_igvn(root());
2425 }
2426 }
2427
2428 //------------------------------uncommon_trap----------------------------------
2429 // Bail out to the interpreter in mid-method. Implemented by calling the
2430 // uncommon_trap blob. This helper function inserts a runtime call with the
2431 // right debug info.
2432 Node* GraphKit::uncommon_trap(int trap_request,
2433 ciKlass* klass, const char* comment,
2434 bool must_throw,
2435 bool keep_exact_action) {
2436 if (failing_internal()) {
2437 stop();
2438 }
2439 if (stopped()) return nullptr; // trap reachable?
2440
2441 // Note: If ProfileTraps is true, and if a deopt. actually
2442 // occurs here, the runtime will make sure an MDO exists. There is
2443 // no need to call method()->ensure_method_data() at this point.
2444
2445 // Set the stack pointer to the right value for reexecution:
2587 *
2588 * @param n node that the type applies to
2589 * @param exact_kls type from profiling
2590 * @param maybe_null did profiling see null?
2591 *
2592 * @return node with improved type
2593 */
2594 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2595 const Type* current_type = _gvn.type(n);
2596 assert(UseTypeSpeculation, "type speculation must be on");
2597
2598 const TypePtr* speculative = current_type->speculative();
2599
2600 // Should the klass from the profile be recorded in the speculative type?
2601 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2602 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2603 const TypeOopPtr* xtype = tklass->as_instance_type();
2604 assert(xtype->klass_is_exact(), "Should be exact");
2605 // Any reason to believe n is not null (from this profiling or a previous one)?
2606 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2607 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2608 // record the new speculative type's depth
2609 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2610 speculative = speculative->with_inline_depth(jvms()->depth());
2611 } else if (current_type->would_improve_ptr(ptr_kind)) {
2612 // Profiling report that null was never seen so we can change the
2613 // speculative type to non null ptr.
2614 if (ptr_kind == ProfileAlwaysNull) {
2615 speculative = TypePtr::NULL_PTR;
2616 } else {
2617 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2618 const TypePtr* ptr = TypePtr::NOTNULL;
2619 if (speculative != nullptr) {
2620 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2621 } else {
2622 speculative = ptr;
2623 }
2624 }
2625 }
2626
2627 if (speculative != current_type->speculative()) {
2628 // Build a type with a speculative type (what we think we know
2629 // about the type but will need a guard when we use it)
2630 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2631 // We're changing the type, we need a new CheckCast node to carry
2632 // the new type. The new type depends on the control: what
2633 // profiling tells us is only valid from here as far as we can
2634 // tell.
2635 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2636 cast = _gvn.transform(cast);
2637 replace_in_map(n, cast);
2638 n = cast;
2639 }
2640
2641 return n;
2642 }
2643
2644 /**
2645 * Record profiling data from receiver profiling at an invoke with the
2646 * type system so that it can propagate it (speculation)
2647 *
2648 * @param n receiver node
2649 *
2650 * @return node with improved type
2651 */
2652 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2653 if (!UseTypeSpeculation) {
2654 return n;
2655 }
2656 ciKlass* exact_kls = profile_has_unique_klass();
2657 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2658 if ((java_bc() == Bytecodes::_checkcast ||
2659 java_bc() == Bytecodes::_instanceof ||
2660 java_bc() == Bytecodes::_aastore) &&
2661 method()->method_data()->is_mature()) {
2662 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2663 if (data != nullptr) {
2664 if (java_bc() == Bytecodes::_aastore) {
2665 ciKlass* array_type = nullptr;
2666 ciKlass* element_type = nullptr;
2667 ProfilePtrKind element_ptr = ProfileMaybeNull;
2668 bool flat_array = true;
2669 bool null_free_array = true;
2670 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2671 exact_kls = element_type;
2672 ptr_kind = element_ptr;
2673 } else {
2674 if (!data->as_BitData()->null_seen()) {
2675 ptr_kind = ProfileNeverNull;
2676 } else {
2677 if (TypeProfileCasts) {
2678 assert(data->is_ReceiverTypeData(), "bad profile data type");
2679 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2680 uint i = 0;
2681 for (; i < call->row_limit(); i++) {
2682 ciKlass* receiver = call->receiver(i);
2683 if (receiver != nullptr) {
2684 break;
2685 }
2686 }
2687 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2688 }
2689 }
2690 }
2691 }
2692 }
2693 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2694 }
2695
2696 /**
2697 * Record profiling data from argument profiling at an invoke with the
2698 * type system so that it can propagate it (speculation)
2699 *
2700 * @param dest_method target method for the call
2701 * @param bc what invoke bytecode is this?
2702 */
2703 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2704 if (!UseTypeSpeculation) {
2705 return;
2706 }
2707 const TypeFunc* tf = TypeFunc::make(dest_method);
2708 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2709 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2710 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2711 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2712 if (is_reference_type(targ->basic_type())) {
2713 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2714 ciKlass* better_type = nullptr;
2715 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2716 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2717 }
2718 i++;
2719 }
2720 }
2721 }
2722
2723 /**
2724 * Record profiling data from parameter profiling at an invoke with
2725 * the type system so that it can propagate it (speculation)
2726 */
2727 void GraphKit::record_profiled_parameters_for_speculation() {
2728 if (!UseTypeSpeculation) {
2729 return;
2730 }
2731 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2851 // The first null ends the list.
2852 Node* parm0, Node* parm1,
2853 Node* parm2, Node* parm3,
2854 Node* parm4, Node* parm5,
2855 Node* parm6, Node* parm7) {
2856 assert(call_addr != nullptr, "must not call null targets");
2857
2858 // Slow-path call
2859 bool is_leaf = !(flags & RC_NO_LEAF);
2860 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2861 if (call_name == nullptr) {
2862 assert(!is_leaf, "must supply name for leaf");
2863 call_name = OptoRuntime::stub_name(call_addr);
2864 }
2865 CallNode* call;
2866 if (!is_leaf) {
2867 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2868 } else if (flags & RC_NO_FP) {
2869 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2870 } else if (flags & RC_VECTOR){
2871 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2872 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2873 } else if (flags & RC_PURE) {
2874 assert(adr_type == nullptr, "pure call does not touch memory");
2875 call = new CallLeafPureNode(call_type, call_addr, call_name);
2876 } else {
2877 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2878 }
2879
2880 // The following is similar to set_edges_for_java_call,
2881 // except that the memory effects of the call are restricted to AliasIdxRaw.
2882
2883 // Slow path call has no side-effects, uses few values
2884 bool wide_in = !(flags & RC_NARROW_MEM);
2885 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2886
2887 Node* prev_mem = nullptr;
2888 if (wide_in) {
2889 prev_mem = set_predefined_input_for_runtime_call(call);
2890 } else {
2891 assert(!wide_out, "narrow in => narrow out");
2892 Node* narrow_mem = memory(adr_type);
2893 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2894 }
2895
2896 // Hook each parm in order. Stop looking at the first null.
2897 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2898 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2899 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2900 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2901 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2902 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2903 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2904 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2905 /* close each nested if ===> */ } } } } } } } }
2906 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2907
2908 if (!is_leaf) {
2909 // Non-leaves can block and take safepoints:
2910 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2911 }
2912 // Non-leaves can throw exceptions:
2913 if (has_io) {
2914 call->set_req(TypeFunc::I_O, i_o());
2915 }
2916
2917 if (flags & RC_UNCOMMON) {
2918 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2919 // (An "if" probability corresponds roughly to an unconditional count.
2920 // Sort of.)
2921 call->set_cnt(PROB_UNLIKELY_MAG(4));
2922 }
2923
2924 Node* c = _gvn.transform(call);
2925 assert(c == call, "cannot disappear");
2926
2934
2935 if (has_io) {
2936 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2937 }
2938 return call;
2939
2940 }
2941
2942 // i2b
2943 Node* GraphKit::sign_extend_byte(Node* in) {
2944 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2945 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2946 }
2947
2948 // i2s
2949 Node* GraphKit::sign_extend_short(Node* in) {
2950 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2951 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2952 }
2953
2954
2955 //------------------------------merge_memory-----------------------------------
2956 // Merge memory from one path into the current memory state.
2957 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2958 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2959 Node* old_slice = mms.force_memory();
2960 Node* new_slice = mms.memory2();
2961 if (old_slice != new_slice) {
2962 PhiNode* phi;
2963 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2964 if (mms.is_empty()) {
2965 // clone base memory Phi's inputs for this memory slice
2966 assert(old_slice == mms.base_memory(), "sanity");
2967 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2968 _gvn.set_type(phi, Type::MEMORY);
2969 for (uint i = 1; i < phi->req(); i++) {
2970 phi->init_req(i, old_slice->in(i));
2971 }
2972 } else {
2973 phi = old_slice->as_Phi(); // Phi was generated already
2974 }
3031 gvn.transform(iff);
3032 if (!bol->is_Con()) gvn.record_for_igvn(iff);
3033 return iff;
3034 }
3035
3036 //-------------------------------gen_subtype_check-----------------------------
3037 // Generate a subtyping check. Takes as input the subtype and supertype.
3038 // Returns 2 values: sets the default control() to the true path and returns
3039 // the false path. Only reads invariant memory; sets no (visible) memory.
3040 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3041 // but that's not exposed to the optimizer. This call also doesn't take in an
3042 // Object; if you wish to check an Object you need to load the Object's class
3043 // prior to coming here.
3044 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3045 ciMethod* method, int bci) {
3046 Compile* C = gvn.C;
3047 if ((*ctrl)->is_top()) {
3048 return C->top();
3049 }
3050
3051 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3052 // For a direct pointer comparison, we need the refined array klass pointer
3053 Node* vm_superklass = superklass;
3054 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3055 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3056 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3057 }
3058
3059 // Fast check for identical types, perhaps identical constants.
3060 // The types can even be identical non-constants, in cases
3061 // involving Array.newInstance, Object.clone, etc.
3062 if (subklass == superklass)
3063 return C->top(); // false path is dead; no test needed.
3064
3065 if (gvn.type(superklass)->singleton()) {
3066 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3067 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3068
3069 // In the common case of an exact superklass, try to fold up the
3070 // test before generating code. You may ask, why not just generate
3071 // the code and then let it fold up? The answer is that the generated
3072 // code will necessarily include null checks, which do not always
3073 // completely fold away. If they are also needless, then they turn
3074 // into a performance loss. Example:
3075 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3076 // Here, the type of 'fa' is often exact, so the store check
3077 // of fa[1]=x will fold up, without testing the nullness of x.
3078 //
3079 // At macro expansion, we would have already folded the SubTypeCheckNode
3080 // being expanded here because we always perform the static sub type
3081 // check in SubTypeCheckNode::sub() regardless of whether
3082 // StressReflectiveCode is set or not. We can therefore skip this
3083 // static check when StressReflectiveCode is on.
3084 switch (C->static_subtype_check(superk, subk)) {
3085 case Compile::SSC_always_false:
3086 {
3087 Node* always_fail = *ctrl;
3088 *ctrl = gvn.C->top();
3089 return always_fail;
3090 }
3091 case Compile::SSC_always_true:
3092 return C->top();
3093 case Compile::SSC_easy_test:
3094 {
3095 // Just do a direct pointer compare and be done.
3096 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3097 *ctrl = gvn.transform(new IfTrueNode(iff));
3098 return gvn.transform(new IfFalseNode(iff));
3099 }
3100 case Compile::SSC_full_test:
3101 break;
3102 default:
3103 ShouldNotReachHere();
3104 }
3105 }
3106
3107 // %%% Possible further optimization: Even if the superklass is not exact,
3108 // if the subklass is the unique subtype of the superklass, the check
3109 // will always succeed. We could leave a dependency behind to ensure this.
3110
3111 // First load the super-klass's check-offset
3112 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3113 Node* m = C->immutable_memory();
3114 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3115 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3116 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3154 gvn.record_for_igvn(r_ok_subtype);
3155
3156 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3157 // SubTypeCheck node
3158 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3159 ciCallProfile profile = method->call_profile_at_bci(bci);
3160 float total_prob = 0;
3161 for (int i = 0; profile.has_receiver(i); ++i) {
3162 float prob = profile.receiver_prob(i);
3163 total_prob += prob;
3164 }
3165 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3166 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3167 for (int i = 0; profile.has_receiver(i); ++i) {
3168 ciKlass* klass = profile.receiver(i);
3169 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3170 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3171 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3172 continue;
3173 }
3174 if (klass_t->isa_aryklassptr()) {
3175 // For a direct pointer comparison, we need the refined array klass pointer
3176 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3177 }
3178 float prob = profile.receiver_prob(i);
3179 ConNode* klass_node = gvn.makecon(klass_t);
3180 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3181 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3182
3183 if (result == Compile::SSC_always_true) {
3184 r_ok_subtype->add_req(iftrue);
3185 } else {
3186 assert(result == Compile::SSC_always_false, "");
3187 r_not_subtype->add_req(iftrue);
3188 }
3189 *ctrl = gvn.transform(new IfFalseNode(iff));
3190 }
3191 }
3192 }
3193
3194 // See if we get an immediate positive hit. Happens roughly 83% of the
3195 // time. Test to see if the value loaded just previously from the subklass
3196 // is exactly the superklass.
3197 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3211 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3212 }
3213 return not_subtype_ctrl;
3214 }
3215
3216 r_ok_subtype->init_req(1, iftrue1);
3217
3218 // Check for immediate negative hit. Happens roughly 11% of the time (which
3219 // is roughly 63% of the remaining cases). Test to see if the loaded
3220 // check-offset points into the subklass display list or the 1-element
3221 // cache. If it points to the display (and NOT the cache) and the display
3222 // missed then it's not a subtype.
3223 Node *cacheoff = gvn.intcon(cacheoff_con);
3224 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3225 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3226 *ctrl = gvn.transform(new IfFalseNode(iff2));
3227
3228 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3229 // No performance impact (too rare) but allows sharing of secondary arrays
3230 // which has some footprint reduction.
3231 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3232 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3233 *ctrl = gvn.transform(new IfFalseNode(iff3));
3234
3235 // -- Roads not taken here: --
3236 // We could also have chosen to perform the self-check at the beginning
3237 // of this code sequence, as the assembler does. This would not pay off
3238 // the same way, since the optimizer, unlike the assembler, can perform
3239 // static type analysis to fold away many successful self-checks.
3240 // Non-foldable self checks work better here in second position, because
3241 // the initial primary superclass check subsumes a self-check for most
3242 // types. An exception would be a secondary type like array-of-interface,
3243 // which does not appear in its own primary supertype display.
3244 // Finally, we could have chosen to move the self-check into the
3245 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3246 // dependent manner. But it is worthwhile to have the check here,
3247 // where it can be perhaps be optimized. The cost in code space is
3248 // small (register compare, branch).
3249
3250 // Now do a linear scan of the secondary super-klass array. Again, no real
3251 // performance impact (too rare) but it's gotta be done.
3252 // Since the code is rarely used, there is no penalty for moving it
3253 // out of line, and it can only improve I-cache density.
3254 // The decision to inline or out-of-line this final check is platform
3255 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3256 Node* psc = gvn.transform(
3257 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3258
3259 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3260 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3261 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3262
3263 // Return false path; set default control to true path.
3264 *ctrl = gvn.transform(r_ok_subtype);
3265 return gvn.transform(r_not_subtype);
3266 }
3267
3268 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3269 const Type* sub_t = _gvn.type(obj_or_subklass);
3270 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3271 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3272 obj_or_subklass = makecon(sub_t);
3273 }
3274 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3275 if (expand_subtype_check) {
3276 MergeMemNode* mem = merged_memory();
3277 Node* ctrl = control();
3278 Node* subklass = obj_or_subklass;
3279 if (!sub_t->isa_klassptr()) {
3280 subklass = load_object_klass(obj_or_subklass);
3281 }
3282
3283 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3284 set_control(ctrl);
3285 return n;
3286 }
3287
3288 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3289 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3290 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3291 set_control(_gvn.transform(new IfTrueNode(iff)));
3292 return _gvn.transform(new IfFalseNode(iff));
3293 }
3294
3295 // Profile-driven exact type check:
3296 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3297 float prob, Node* *casted_receiver) {
3298 assert(!klass->is_interface(), "no exact type check on interfaces");
3299 Node* fail = top();
3300 const Type* rec_t = _gvn.type(receiver);
3301 if (rec_t->is_inlinetypeptr()) {
3302 if (klass->equals(rec_t->inline_klass())) {
3303 (*casted_receiver) = receiver; // Always passes
3304 } else {
3305 (*casted_receiver) = top(); // Always fails
3306 fail = control();
3307 set_control(top());
3308 }
3309 return fail;
3310 }
3311 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3312 if (tklass->isa_aryklassptr()) {
3313 // For a direct pointer comparison, we need the refined array klass pointer
3314 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3315 }
3316 Node* recv_klass = load_object_klass(receiver);
3317 fail = type_check(recv_klass, tklass, prob);
3318
3319 if (!stopped()) {
3320 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3321 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3322 assert(recv_xtype->klass_is_exact(), "");
3323
3324 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3325 // Subsume downstream occurrences of receiver with a cast to
3326 // recv_xtype, since now we know what the type will be.
3327 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3328 Node* res = _gvn.transform(cast);
3329 if (recv_xtype->is_inlinetypeptr()) {
3330 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3331 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3332 }
3333 (*casted_receiver) = res;
3334 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3335 // (User must make the replace_in_map call.)
3336 }
3337 }
3338
3339 return fail;
3340 }
3341
3342 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3343 float prob) {
3344 Node* want_klass = makecon(tklass);
3345 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3346 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3347 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3348 set_control(_gvn.transform(new IfTrueNode (iff)));
3349 Node* fail = _gvn.transform(new IfFalseNode(iff));
3350 return fail;
3351 }
3352
3353 //------------------------------subtype_check_receiver-------------------------
3354 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3355 Node** casted_receiver) {
3356 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3357 Node* want_klass = makecon(tklass);
3358
3359 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3360
3361 // Ignore interface type information until interface types are properly tracked.
3362 if (!stopped() && !klass->is_interface()) {
3363 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3364 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3365 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3366 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3367 if (recv_type->is_inlinetypeptr()) {
3368 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3369 }
3370 (*casted_receiver) = cast;
3371 }
3372 }
3373
3374 return slow_ctl;
3375 }
3376
3377 //------------------------------seems_never_null-------------------------------
3378 // Use null_seen information if it is available from the profile.
3379 // If we see an unexpected null at a type check we record it and force a
3380 // recompile; the offending check will be recompiled to handle nulls.
3381 // If we see several offending BCIs, then all checks in the
3382 // method will be recompiled.
3383 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3384 speculating = !_gvn.type(obj)->speculative_maybe_null();
3385 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3386 if (UncommonNullCast // Cutout for this technique
3387 && obj != null() // And not the -Xcomp stupid case?
3388 && !too_many_traps(reason)
3389 ) {
3390 if (speculating) {
3459
3460 //------------------------maybe_cast_profiled_receiver-------------------------
3461 // If the profile has seen exactly one type, narrow to exactly that type.
3462 // Subsequent type checks will always fold up.
3463 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3464 const TypeKlassPtr* require_klass,
3465 ciKlass* spec_klass,
3466 bool safe_for_replace) {
3467 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3468
3469 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3470
3471 // Make sure we haven't already deoptimized from this tactic.
3472 if (too_many_traps_or_recompiles(reason))
3473 return nullptr;
3474
3475 // (No, this isn't a call, but it's enough like a virtual call
3476 // to use the same ciMethod accessor to get the profile info...)
3477 // If we have a speculative type use it instead of profiling (which
3478 // may not help us)
3479 ciKlass* exact_kls = spec_klass;
3480 if (exact_kls == nullptr) {
3481 if (java_bc() == Bytecodes::_aastore) {
3482 ciKlass* array_type = nullptr;
3483 ciKlass* element_type = nullptr;
3484 ProfilePtrKind element_ptr = ProfileMaybeNull;
3485 bool flat_array = true;
3486 bool null_free_array = true;
3487 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3488 exact_kls = element_type;
3489 } else {
3490 exact_kls = profile_has_unique_klass();
3491 }
3492 }
3493 if (exact_kls != nullptr) {// no cast failures here
3494 if (require_klass == nullptr ||
3495 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3496 // If we narrow the type to match what the type profile sees or
3497 // the speculative type, we can then remove the rest of the
3498 // cast.
3499 // This is a win, even if the exact_kls is very specific,
3500 // because downstream operations, such as method calls,
3501 // will often benefit from the sharper type.
3502 Node* exact_obj = not_null_obj; // will get updated in place...
3503 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3504 &exact_obj);
3505 { PreserveJVMState pjvms(this);
3506 set_control(slow_ctl);
3507 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3508 }
3509 if (safe_for_replace) {
3510 replace_in_map(not_null_obj, exact_obj);
3511 }
3512 return exact_obj;
3602 // If not_null_obj is dead, only null-path is taken
3603 if (stopped()) { // Doing instance-of on a null?
3604 set_control(null_ctl);
3605 return intcon(0);
3606 }
3607 region->init_req(_null_path, null_ctl);
3608 phi ->init_req(_null_path, intcon(0)); // Set null path value
3609 if (null_ctl == top()) {
3610 // Do this eagerly, so that pattern matches like is_diamond_phi
3611 // will work even during parsing.
3612 assert(_null_path == PATH_LIMIT-1, "delete last");
3613 region->del_req(_null_path);
3614 phi ->del_req(_null_path);
3615 }
3616
3617 // Do we know the type check always succeed?
3618 bool known_statically = false;
3619 if (_gvn.type(superklass)->singleton()) {
3620 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3621 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3622 if (subk != nullptr && subk->is_loaded()) {
3623 int static_res = C->static_subtype_check(superk, subk);
3624 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3625 }
3626 }
3627
3628 if (!known_statically) {
3629 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3630 // We may not have profiling here or it may not help us. If we
3631 // have a speculative type use it to perform an exact cast.
3632 ciKlass* spec_obj_type = obj_type->speculative_type();
3633 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3634 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3635 if (stopped()) { // Profile disagrees with this path.
3636 set_control(null_ctl); // Null is the only remaining possibility.
3637 return intcon(0);
3638 }
3639 if (cast_obj != nullptr) {
3640 not_null_obj = cast_obj;
3641 }
3642 }
3658 record_for_igvn(region);
3659
3660 // If we know the type check always succeeds then we don't use the
3661 // profiling data at this bytecode. Don't lose it, feed it to the
3662 // type system as a speculative type.
3663 if (safe_for_replace) {
3664 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3665 replace_in_map(obj, casted_obj);
3666 }
3667
3668 return _gvn.transform(phi);
3669 }
3670
3671 //-------------------------------gen_checkcast---------------------------------
3672 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3673 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3674 // uncommon-trap paths work. Adjust stack after this call.
3675 // If failure_control is supplied and not null, it is filled in with
3676 // the control edge for the cast failure. Otherwise, an appropriate
3677 // uncommon trap or exception is thrown.
3678 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3679 // when the current map for the success path is updated with information only present
3680 // on the success path and not the cast failure path. The newly cloned map should then be
3681 // used to emit the uncommon trap in the caller.
3682 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3683 assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3684 "failure_control must be set when new_failure_map is used");
3685 kill_dead_locals(); // Benefit all the uncommon traps
3686 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3687 const Type* obj_type = _gvn.type(obj);
3688
3689 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3690 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3691 bool safe_for_replace = (failure_control == nullptr);
3692 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3693
3694 // Fast cutout: Check the case that the cast is vacuously true.
3695 // This detects the common cases where the test will short-circuit
3696 // away completely. We do this before we perform the null check,
3697 // because if the test is going to turn into zero code, we don't
3698 // want a residual null check left around. (Causes a slowdown,
3699 // for example, in some objArray manipulations, such as a[i]=a[j].)
3700 if (improved_klass_ptr_type->singleton()) {
3701 const TypeKlassPtr* kptr = nullptr;
3702 if (obj_type->isa_oop_ptr()) {
3703 kptr = obj_type->is_oopptr()->as_klass_type();
3704 } else if (obj->is_InlineType()) {
3705 ciInlineKlass* vk = obj_type->inline_klass();
3706 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3707 }
3708
3709 if (kptr != nullptr) {
3710 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3711 case Compile::SSC_always_true:
3712 // If we know the type check always succeed then we don't use
3713 // the profiling data at this bytecode. Don't lose it, feed it
3714 // to the type system as a speculative type.
3715 obj = record_profiled_receiver_for_speculation(obj);
3716 if (null_free) {
3717 assert(safe_for_replace, "must be");
3718 obj = null_check(obj);
3719 }
3720 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3721 return obj;
3722 case Compile::SSC_always_false:
3723 if (null_free) {
3724 assert(safe_for_replace, "must be");
3725 obj = null_check(obj);
3726 }
3727 // It needs a null check because a null will *pass* the cast check.
3728 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3729 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3730 Deoptimization::DeoptReason reason = is_aastore ?
3731 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3732 builtin_throw(reason);
3733 return top();
3734 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3735 return null_assert(obj);
3736 }
3737 break; // Fall through to full check
3738 default:
3739 break;
3740 }
3741 }
3742 }
3743
3744 ciProfileData* data = nullptr;
3745 if (failure_control == nullptr) { // use MDO in regular case only
3746 assert(java_bc() == Bytecodes::_aastore ||
3747 java_bc() == Bytecodes::_checkcast,
3748 "interpreter profiles type checks only for these BCs");
3749 if (method()->method_data()->is_mature()) {
3750 data = method()->method_data()->bci_to_data(bci());
3751 }
3752 }
3753
3754 // Make the merge point
3755 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3756 RegionNode* region = new RegionNode(PATH_LIMIT);
3757 Node* phi = new PhiNode(region, toop);
3758 _gvn.set_type(region, Type::CONTROL);
3759 _gvn.set_type(phi, toop);
3760
3761 C->set_has_split_ifs(true); // Has chance for split-if optimization
3762
3763 // Use null-cast information if it is available
3764 bool speculative_not_null = false;
3765 bool never_see_null = ((failure_control == nullptr) // regular case only
3766 && seems_never_null(obj, data, speculative_not_null));
3767
3768 if (obj->is_InlineType()) {
3769 // Re-execute if buffering during triggers deoptimization
3770 PreserveReexecuteState preexecs(this);
3771 jvms()->set_should_reexecute(true);
3772 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3773 }
3774
3775 // Null check; get casted pointer; set region slot 3
3776 Node* null_ctl = top();
3777 Node* not_null_obj = nullptr;
3778 if (null_free) {
3779 assert(safe_for_replace, "must be");
3780 not_null_obj = null_check(obj);
3781 } else {
3782 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3783 }
3784
3785 // If not_null_obj is dead, only null-path is taken
3786 if (stopped()) { // Doing instance-of on a null?
3787 set_control(null_ctl);
3788 if (toop->is_inlinetypeptr()) {
3789 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3790 }
3791 return null();
3792 }
3793 region->init_req(_null_path, null_ctl);
3794 phi ->init_req(_null_path, null()); // Set null path value
3795 if (null_ctl == top()) {
3796 // Do this eagerly, so that pattern matches like is_diamond_phi
3797 // will work even during parsing.
3798 assert(_null_path == PATH_LIMIT-1, "delete last");
3799 region->del_req(_null_path);
3800 phi ->del_req(_null_path);
3801 }
3802
3803 Node* cast_obj = nullptr;
3804 if (improved_klass_ptr_type->klass_is_exact()) {
3805 // The following optimization tries to statically cast the speculative type of the object
3806 // (for example obtained during profiling) to the type of the superklass and then do a
3807 // dynamic check that the type of the object is what we expect. To work correctly
3808 // for checkcast and aastore the type of superklass should be exact.
3809 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3810 // We may not have profiling here or it may not help us. If we have
3811 // a speculative type use it to perform an exact cast.
3812 ciKlass* spec_obj_type = obj_type->speculative_type();
3813 if (spec_obj_type != nullptr || data != nullptr) {
3814 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3815 if (cast_obj != nullptr) {
3816 if (failure_control != nullptr) // failure is now impossible
3817 (*failure_control) = top();
3818 // adjust the type of the phi to the exact klass:
3819 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3820 }
3821 }
3822 }
3823
3824 if (cast_obj == nullptr) {
3825 // Generate the subtype check
3826 Node* improved_superklass = superklass;
3827 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3828 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3829 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3830 // Additionally, the benefit would only be minor in non-constant cases.
3831 improved_superklass = makecon(improved_klass_ptr_type);
3832 }
3833 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3834 // Plug in success path into the merge
3835 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3836 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3837 if (failure_control == nullptr) {
3838 if (not_subtype_ctrl != top()) { // If failure is possible
3839 PreserveJVMState pjvms(this);
3840 set_control(not_subtype_ctrl);
3841 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3842 Deoptimization::DeoptReason reason = is_aastore ?
3843 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3844 builtin_throw(reason);
3845 }
3846 } else {
3847 (*failure_control) = not_subtype_ctrl;
3848 }
3849 }
3850
3851 region->init_req(_obj_path, control());
3852 phi ->init_req(_obj_path, cast_obj);
3853
3854 // A merge of null or Casted-NotNull obj
3855 Node* res = _gvn.transform(phi);
3856
3857 // Note I do NOT always 'replace_in_map(obj,result)' here.
3858 // if( tk->klass()->can_be_primary_super() )
3859 // This means that if I successfully store an Object into an array-of-String
3860 // I 'forget' that the Object is really now known to be a String. I have to
3861 // do this because we don't have true union types for interfaces - if I store
3862 // a Baz into an array-of-Interface and then tell the optimizer it's an
3863 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3864 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3865 // replace_in_map( obj, res );
3866
3867 // Return final merged results
3868 set_control( _gvn.transform(region) );
3869 record_for_igvn(region);
3870
3871 bool not_inline = !toop->can_be_inline_type();
3872 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3873 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3874 // Check if obj has been loaded from an array
3875 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3876 Node* array = nullptr;
3877 if (obj->isa_Load()) {
3878 Node* address = obj->in(MemNode::Address);
3879 if (address->isa_AddP()) {
3880 array = address->as_AddP()->in(AddPNode::Base);
3881 }
3882 } else if (obj->is_Phi()) {
3883 Node* region = obj->in(0);
3884 // TODO make this more robust (see JDK-8231346)
3885 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3886 IfNode* iff = region->in(2)->in(0)->isa_If();
3887 if (iff != nullptr) {
3888 iff->is_flat_array_check(&_gvn, &array);
3889 }
3890 }
3891 }
3892 if (array != nullptr) {
3893 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3894 if (ary_t != nullptr) {
3895 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3896 // Casting array element to a non-inline-type, mark array as not null-free.
3897 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3898 if (new_cast_failure_map != nullptr) {
3899 // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3900 // create the cast failure path in the caller without wrongly making the cast node live there.
3901 *new_cast_failure_map = clone_map();
3902 }
3903 replace_in_map(array, cast);
3904 array = cast;
3905 }
3906 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3907 // Casting array element to a non-flat-in-array type, mark array as not flat.
3908 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3909 if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3910 // Same as above.
3911 *new_cast_failure_map = clone_map();
3912 }
3913 replace_in_map(array, cast);
3914 array = cast;
3915 }
3916 }
3917 }
3918 }
3919
3920 if (!stopped() && !res->is_InlineType()) {
3921 res = record_profiled_receiver_for_speculation(res);
3922 if (toop->is_inlinetypeptr() && !maybe_larval) {
3923 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3924 res = vt;
3925 if (safe_for_replace) {
3926 replace_in_map(obj, vt);
3927 replace_in_map(not_null_obj, vt);
3928 replace_in_map(res, vt);
3929 }
3930 }
3931 }
3932 return res;
3933 }
3934
3935 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3936 // Load markword
3937 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3938 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3939 if (check_lock && !UseCompactObjectHeaders) {
3940 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3941 // Check if obj is locked
3942 Node* locked_bit = MakeConX(markWord::unlocked_value);
3943 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3944 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3945 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3946 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3947 _gvn.transform(iff);
3948 Node* locked_region = new RegionNode(3);
3949 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3950
3951 // Unlocked: Use bits from mark word
3952 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3953 mark_phi->init_req(1, mark);
3954
3955 // Locked: Load prototype header from klass
3956 set_control(_gvn.transform(new IfFalseNode(iff)));
3957 // Make loads control dependent to make sure they are only executed if array is locked
3958 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3959 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3960 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3961 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3962
3963 locked_region->init_req(2, control());
3964 mark_phi->init_req(2, proto);
3965 set_control(_gvn.transform(locked_region));
3966 record_for_igvn(locked_region);
3967
3968 mark = mark_phi;
3969 }
3970
3971 // Now check if mark word bits are set
3972 Node* mask = MakeConX(mask_val);
3973 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3974 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3975 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3976 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3977 }
3978
3979 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3980 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3981 }
3982
3983 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3984 // We can't use immutable memory here because the mark word is mutable.
3985 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3986 // check is moved out of loops (mainly to enable loop unswitching).
3987 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3988 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3989 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3990 }
3991
3992 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3993 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3994 }
3995
3996 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3997 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3998
3999 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
4000 if (!vk->has_null_free_non_atomic_layout()) {
4001 return intcon(1); // Always atomic
4002 } else if (!vk->has_null_free_atomic_layout()) {
4003 return intcon(0); // Never atomic
4004 }
4005
4006 Node* array_klass = load_object_klass(array);
4007 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4008 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4009 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4010 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
4011 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4012 }
4013
4014 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4015 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4016 RegionNode* region = new RegionNode(3);
4017 Node* null_ctl = top();
4018 null_check_oop(val, &null_ctl);
4019 if (null_ctl != top()) {
4020 PreserveJVMState pjvms(this);
4021 set_control(null_ctl);
4022 {
4023 // Deoptimize if null-free array
4024 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4025 inc_sp(nargs);
4026 uncommon_trap(Deoptimization::Reason_null_check,
4027 Deoptimization::Action_none);
4028 }
4029 region->init_req(1, control());
4030 }
4031 region->init_req(2, control());
4032 set_control(_gvn.transform(region));
4033 record_for_igvn(region);
4034 if (_gvn.type(val) == TypePtr::NULL_PTR) {
4035 // Since we were just successfully storing null, the array can't be null free.
4036 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4037 ary_t = ary_t->cast_to_not_null_free();
4038 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4039 if (safe_for_replace) {
4040 replace_in_map(ary, cast);
4041 }
4042 ary = cast;
4043 }
4044 return ary;
4045 }
4046
4047 //------------------------------next_monitor-----------------------------------
4048 // What number should be given to the next monitor?
4049 int GraphKit::next_monitor() {
4050 int current = jvms()->monitor_depth()* C->sync_stack_slots();
4051 int next = current + C->sync_stack_slots();
4052 // Keep the toplevel high water mark current:
4053 if (C->fixed_slots() < next) C->set_fixed_slots(next);
4054 return current;
4055 }
4056
4057 //------------------------------insert_mem_bar---------------------------------
4058 // Memory barrier to avoid floating things around
4059 // The membar serves as a pinch point between both control and all memory slices.
4060 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4061 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4062 mb->init_req(TypeFunc::Control, control());
4063 mb->init_req(TypeFunc::Memory, reset_memory());
4064 Node* membar = _gvn.transform(mb);
4158 lock->create_lock_counter(map()->jvms());
4159 increment_counter(lock->counter()->addr());
4160 }
4161 #endif
4162
4163 return flock;
4164 }
4165
4166
4167 //------------------------------shared_unlock----------------------------------
4168 // Emit unlocking code.
4169 void GraphKit::shared_unlock(Node* box, Node* obj) {
4170 // bci is either a monitorenter bc or InvocationEntryBci
4171 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4172 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4173
4174 if (stopped()) { // Dead monitor?
4175 map()->pop_monitor(); // Kill monitor from debug info
4176 return;
4177 }
4178 assert(!obj->is_InlineType(), "should not unlock on inline type");
4179
4180 // Memory barrier to avoid floating things down past the locked region
4181 insert_mem_bar(Op_MemBarReleaseLock);
4182
4183 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4184 UnlockNode *unlock = new UnlockNode(C, tf);
4185 #ifdef ASSERT
4186 unlock->set_dbg_jvms(sync_jvms());
4187 #endif
4188 uint raw_idx = Compile::AliasIdxRaw;
4189 unlock->init_req( TypeFunc::Control, control() );
4190 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4191 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4192 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4193 unlock->init_req( TypeFunc::ReturnAdr, top() );
4194
4195 unlock->init_req(TypeFunc::Parms + 0, obj);
4196 unlock->init_req(TypeFunc::Parms + 1, box);
4197 unlock = _gvn.transform(unlock)->as_Unlock();
4198
4199 Node* mem = reset_memory();
4200
4201 // unlock has no side-effects, sets few values
4202 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4203
4204 // Kill monitor from debug info
4205 map()->pop_monitor( );
4206 }
4207
4208 //-------------------------------get_layout_helper-----------------------------
4209 // If the given klass is a constant or known to be an array,
4210 // fetch the constant layout helper value into constant_value
4211 // and return null. Otherwise, load the non-constant
4212 // layout helper value, and return the node which represents it.
4213 // This two-faced routine is useful because allocation sites
4214 // almost always feature constant types.
4215 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4216 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4217 if (!StressReflectiveCode && klass_t != nullptr) {
4218 bool xklass = klass_t->klass_is_exact();
4219 bool can_be_flat = false;
4220 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4221 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4222 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4223 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4224 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4225 }
4226 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4227 jint lhelper;
4228 if (klass_t->is_flat()) {
4229 lhelper = ary_type->flat_layout_helper();
4230 } else if (klass_t->isa_aryklassptr()) {
4231 BasicType elem = ary_type->elem()->array_element_basic_type();
4232 if (is_reference_type(elem, true)) {
4233 elem = T_OBJECT;
4234 }
4235 lhelper = Klass::array_layout_helper(elem);
4236 } else {
4237 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4238 }
4239 if (lhelper != Klass::_lh_neutral_value) {
4240 constant_value = lhelper;
4241 return (Node*) nullptr;
4242 }
4243 }
4244 }
4245 constant_value = Klass::_lh_neutral_value; // put in a known value
4246 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4247 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4248 }
4249
4250 // We just put in an allocate/initialize with a big raw-memory effect.
4251 // Hook selected additional alias categories on the initialization.
4252 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4253 MergeMemNode* init_in_merge,
4254 Node* init_out_raw) {
4255 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4256 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4257
4258 Node* prevmem = kit.memory(alias_idx);
4259 init_in_merge->set_memory_at(alias_idx, prevmem);
4260 if (init_out_raw != nullptr) {
4261 kit.set_memory(init_out_raw, alias_idx);
4262 }
4263 }
4264
4265 //---------------------------set_output_for_allocation-------------------------
4266 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4267 const TypeOopPtr* oop_type,
4268 bool deoptimize_on_exception) {
4269 int rawidx = Compile::AliasIdxRaw;
4270 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4271 add_safepoint_edges(alloc);
4272 Node* allocx = _gvn.transform(alloc);
4273 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4274 // create memory projection for i_o
4275 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4276 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4277
4278 // create a memory projection as for the normal control path
4279 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4280 set_memory(malloc, rawidx);
4281
4282 // a normal slow-call doesn't change i_o, but an allocation does
4283 // we create a separate i_o projection for the normal control path
4284 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4285 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4286
4287 // put in an initialization barrier
4288 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4289 rawoop)->as_Initialize();
4290 assert(alloc->initialization() == init, "2-way macro link must work");
4291 assert(init ->allocation() == alloc, "2-way macro link must work");
4292 {
4293 // Extract memory strands which may participate in the new object's
4294 // initialization, and source them from the new InitializeNode.
4295 // This will allow us to observe initializations when they occur,
4296 // and link them properly (as a group) to the InitializeNode.
4297 assert(init->in(InitializeNode::Memory) == malloc, "");
4298 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4299 init->set_req(InitializeNode::Memory, minit_in);
4300 record_for_igvn(minit_in); // fold it up later, if possible
4301 _gvn.set_type(minit_in, Type::MEMORY);
4302 Node* minit_out = memory(rawidx);
4303 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4304 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4305 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4306 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4307 // multiple projections as a result.
4308 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4309 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4310 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4311 if (oop_type->isa_aryptr()) {
4312 // Initially all flat array accesses share a single slice
4313 // but that changes after parsing. Prepare the memory graph so
4314 // it can optimize flat array accesses properly once they
4315 // don't share a single slice.
4316 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4317 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4318 int elemidx = C->get_alias_index(telemref);
4319 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4320 if (alias_adr_type->is_flat()) {
4321 C->set_flat_accesses();
4322 }
4323 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4324 } else if (oop_type->isa_instptr()) {
4325 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4326 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4327 ciField* field = ik->nonstatic_field_at(i);
4328 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4329 continue; // do not bother to track really large numbers of fields
4330 // Find (or create) the alias category for this field:
4331 int fieldidx = C->alias_type(field)->index();
4332 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4333 }
4334 }
4335 }
4336
4337 // Cast raw oop to the real thing...
4338 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4339 javaoop = _gvn.transform(javaoop);
4340 C->set_recent_alloc(control(), javaoop);
4341 assert(just_allocated_object(control()) == javaoop, "just allocated");
4342
4343 #ifdef ASSERT
4355 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4356 }
4357 }
4358 #endif //ASSERT
4359
4360 return javaoop;
4361 }
4362
4363 //---------------------------new_instance--------------------------------------
4364 // This routine takes a klass_node which may be constant (for a static type)
4365 // or may be non-constant (for reflective code). It will work equally well
4366 // for either, and the graph will fold nicely if the optimizer later reduces
4367 // the type to a constant.
4368 // The optional arguments are for specialized use by intrinsics:
4369 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4370 // - If 'return_size_val', report the total object size to the caller.
4371 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4372 Node* GraphKit::new_instance(Node* klass_node,
4373 Node* extra_slow_test,
4374 Node* *return_size_val,
4375 bool deoptimize_on_exception,
4376 InlineTypeNode* inline_type_node) {
4377 // Compute size in doublewords
4378 // The size is always an integral number of doublewords, represented
4379 // as a positive bytewise size stored in the klass's layout_helper.
4380 // The layout_helper also encodes (in a low bit) the need for a slow path.
4381 jint layout_con = Klass::_lh_neutral_value;
4382 Node* layout_val = get_layout_helper(klass_node, layout_con);
4383 bool layout_is_con = (layout_val == nullptr);
4384
4385 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4386 // Generate the initial go-slow test. It's either ALWAYS (return a
4387 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4388 // case) a computed value derived from the layout_helper.
4389 Node* initial_slow_test = nullptr;
4390 if (layout_is_con) {
4391 assert(!StressReflectiveCode, "stress mode does not use these paths");
4392 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4393 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4394 } else { // reflective case
4395 // This reflective path is used by Unsafe.allocateInstance.
4396 // (It may be stress-tested by specifying StressReflectiveCode.)
4397 // Basically, we want to get into the VM is there's an illegal argument.
4398 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4399 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4400 if (extra_slow_test != intcon(0)) {
4401 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4402 }
4403 // (Macro-expander will further convert this to a Bool, if necessary.)
4414
4415 // Clear the low bits to extract layout_helper_size_in_bytes:
4416 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4417 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4418 size = _gvn.transform( new AndXNode(size, mask) );
4419 }
4420 if (return_size_val != nullptr) {
4421 (*return_size_val) = size;
4422 }
4423
4424 // This is a precise notnull oop of the klass.
4425 // (Actually, it need not be precise if this is a reflective allocation.)
4426 // It's what we cast the result to.
4427 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4428 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4429 const TypeOopPtr* oop_type = tklass->as_instance_type();
4430
4431 // Now generate allocation code
4432
4433 // The entire memory state is needed for slow path of the allocation
4434 // since GC and deoptimization can happen.
4435 Node *mem = reset_memory();
4436 set_all_memory(mem); // Create new memory state
4437
4438 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4439 control(), mem, i_o(),
4440 size, klass_node,
4441 initial_slow_test, inline_type_node);
4442
4443 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4444 }
4445
4446 //-------------------------------new_array-------------------------------------
4447 // helper for newarray and anewarray
4448 // The 'length' parameter is (obviously) the length of the array.
4449 // The optional arguments are for specialized use by intrinsics:
4450 // - If 'return_size_val', report the non-padded array size (sum of header size
4451 // and array body) to the caller.
4452 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4453 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4454 Node* length, // number of array elements
4455 int nargs, // number of arguments to push back for uncommon trap
4456 Node* *return_size_val,
4457 bool deoptimize_on_exception,
4458 Node* init_val) {
4459 jint layout_con = Klass::_lh_neutral_value;
4460 Node* layout_val = get_layout_helper(klass_node, layout_con);
4461 bool layout_is_con = (layout_val == nullptr);
4462
4463 if (!layout_is_con && !StressReflectiveCode &&
4464 !too_many_traps(Deoptimization::Reason_class_check)) {
4465 // This is a reflective array creation site.
4466 // Optimistically assume that it is a subtype of Object[],
4467 // so that we can fold up all the address arithmetic.
4468 layout_con = Klass::array_layout_helper(T_OBJECT);
4469 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4470 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4471 { BuildCutout unless(this, bol_lh, PROB_MAX);
4472 inc_sp(nargs);
4473 uncommon_trap(Deoptimization::Reason_class_check,
4474 Deoptimization::Action_maybe_recompile);
4475 }
4476 layout_val = nullptr;
4477 layout_is_con = true;
4478 }
4479
4480 // Generate the initial go-slow test. Make sure we do not overflow
4481 // if length is huge (near 2Gig) or negative! We do not need
4482 // exact double-words here, just a close approximation of needed
4483 // double-words. We can't add any offset or rounding bits, lest we
4484 // take a size -1 of bytes and make it positive. Use an unsigned
4485 // compare, so negative sizes look hugely positive.
4486 int fast_size_limit = FastAllocateSizeLimit;
4487 if (layout_is_con) {
4488 assert(!StressReflectiveCode, "stress mode does not use these paths");
4489 // Increase the size limit if we have exact knowledge of array type.
4490 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4491 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4492 }
4493
4494 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4495 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4496
4497 // --- Size Computation ---
4498 // array_size = round_to_heap(array_header + (length << elem_shift));
4499 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4500 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4501 // The rounding mask is strength-reduced, if possible.
4502 int round_mask = MinObjAlignmentInBytes - 1;
4503 Node* header_size = nullptr;
4504 // (T_BYTE has the weakest alignment and size restrictions...)
4505 if (layout_is_con) {
4506 int hsize = Klass::layout_helper_header_size(layout_con);
4507 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4508 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4509 if ((round_mask & ~right_n_bits(eshift)) == 0)
4510 round_mask = 0; // strength-reduce it if it goes away completely
4511 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4512 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4513 assert(header_size_min <= hsize, "generic minimum is smallest");
4514 header_size = intcon(hsize);
4515 } else {
4516 Node* hss = intcon(Klass::_lh_header_size_shift);
4517 Node* hsm = intcon(Klass::_lh_header_size_mask);
4518 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4519 header_size = _gvn.transform(new AndINode(header_size, hsm));
4520 }
4521
4522 Node* elem_shift = nullptr;
4523 if (layout_is_con) {
4524 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4525 if (eshift != 0)
4526 elem_shift = intcon(eshift);
4527 } else {
4528 // There is no need to mask or shift this value.
4529 // The semantics of LShiftINode include an implicit mask to 0x1F.
4530 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4531 elem_shift = layout_val;
4580 }
4581 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4582
4583 if (return_size_val != nullptr) {
4584 // This is the size
4585 (*return_size_val) = non_rounded_size;
4586 }
4587
4588 Node* size = non_rounded_size;
4589 if (round_mask != 0) {
4590 Node* mask1 = MakeConX(round_mask);
4591 size = _gvn.transform(new AddXNode(size, mask1));
4592 Node* mask2 = MakeConX(~round_mask);
4593 size = _gvn.transform(new AndXNode(size, mask2));
4594 }
4595 // else if round_mask == 0, the size computation is self-rounding
4596
4597 // Now generate allocation code
4598
4599 // The entire memory state is needed for slow path of the allocation
4600 // since GC and deoptimization can happen.
4601 Node *mem = reset_memory();
4602 set_all_memory(mem); // Create new memory state
4603
4604 if (initial_slow_test->is_Bool()) {
4605 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4606 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4607 }
4608
4609 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4610 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4611
4612 Node* raw_init_value = nullptr;
4613 if (init_val != nullptr) {
4614 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4615 if (ary_type->is_flat()) {
4616 initial_slow_test = intcon(1);
4617 }
4618
4619 if (UseCompressedOops) {
4620 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4621 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4622 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4623 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4624 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4625 } else {
4626 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4627 }
4628 }
4629
4630 Node* valid_length_test = _gvn.intcon(1);
4631 if (ary_type->isa_aryptr()) {
4632 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4633 jint max = TypeAryPtr::max_array_length(bt);
4634 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4635 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4636 }
4637
4638 // Create the AllocateArrayNode and its result projections
4639 AllocateArrayNode* alloc
4640 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4641 control(), mem, i_o(),
4642 size, klass_node,
4643 initial_slow_test,
4644 length, valid_length_test,
4645 init_val, raw_init_value);
4646 // Cast to correct type. Note that the klass_node may be constant or not,
4647 // and in the latter case the actual array type will be inexact also.
4648 // (This happens via a non-constant argument to inline_native_newArray.)
4649 // In any case, the value of klass_node provides the desired array type.
4650 const TypeInt* length_type = _gvn.find_int_type(length);
4651 if (ary_type->isa_aryptr() && length_type != nullptr) {
4652 // Try to get a better type than POS for the size
4653 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4654 }
4655
4656 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4657
4658 array_ideal_length(alloc, ary_type, true);
4659 return javaoop;
4660 }
4661
4662 // The following "Ideal_foo" functions are placed here because they recognize
4663 // the graph shapes created by the functions immediately above.
4664
4665 //---------------------------Ideal_allocation----------------------------------
4760 void GraphKit::add_parse_predicates(int nargs) {
4761 if (ShortRunningLongLoop) {
4762 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4763 // walking up from the loop.
4764 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4765 }
4766 if (UseLoopPredicate) {
4767 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4768 if (UseProfiledLoopPredicate) {
4769 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4770 }
4771 }
4772 if (UseAutoVectorizationPredicate) {
4773 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4774 }
4775 // Loop Limit Check Predicate should be near the loop.
4776 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4777 }
4778
4779 void GraphKit::sync_kit(IdealKit& ideal) {
4780 reset_memory();
4781 set_all_memory(ideal.merged_memory());
4782 set_i_o(ideal.i_o());
4783 set_control(ideal.ctrl());
4784 }
4785
4786 void GraphKit::final_sync(IdealKit& ideal) {
4787 // Final sync IdealKit and graphKit.
4788 sync_kit(ideal);
4789 }
4790
4791 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4792 Node* len = load_array_length(load_String_value(str, set_ctrl));
4793 Node* coder = load_String_coder(str, set_ctrl);
4794 // Divide length by 2 if coder is UTF16
4795 return _gvn.transform(new RShiftINode(len, coder));
4796 }
4797
4798 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4799 int value_offset = java_lang_String::value_offset();
4800 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4801 false, nullptr, Type::Offset(0));
4802 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4803 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4804 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4805 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4806 Node* p = basic_plus_adr(str, str, value_offset);
4807 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4808 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4809 return load;
4810 }
4811
4812 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4813 if (!CompactStrings) {
4814 return intcon(java_lang_String::CODER_UTF16);
4815 }
4816 int coder_offset = java_lang_String::coder_offset();
4817 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4818 false, nullptr, Type::Offset(0));
4819 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4820
4821 Node* p = basic_plus_adr(str, str, coder_offset);
4822 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4823 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4824 return load;
4825 }
4826
4827 void GraphKit::store_String_value(Node* str, Node* value) {
4828 int value_offset = java_lang_String::value_offset();
4829 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4830 false, nullptr, Type::Offset(0));
4831 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4832
4833 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4834 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4835 }
4836
4837 void GraphKit::store_String_coder(Node* str, Node* value) {
4838 int coder_offset = java_lang_String::coder_offset();
4839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4840 false, nullptr, Type::Offset(0));
4841 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4842
4843 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4844 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4845 }
4846
4847 // Capture src and dst memory state with a MergeMemNode
4848 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4849 if (src_type == dst_type) {
4850 // Types are equal, we don't need a MergeMemNode
4851 return memory(src_type);
4852 }
4853 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4854 record_for_igvn(merge); // fold it up later, if possible
4855 int src_idx = C->get_alias_index(src_type);
4856 int dst_idx = C->get_alias_index(dst_type);
4857 merge->set_memory_at(src_idx, memory(src_idx));
4858 merge->set_memory_at(dst_idx, memory(dst_idx));
4859 return merge;
4860 }
4933 i_char->init_req(2, AddI(i_char, intcon(2)));
4934
4935 set_control(IfFalse(iff));
4936 set_memory(st, TypeAryPtr::BYTES);
4937 }
4938
4939 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4940 if (!field->is_constant()) {
4941 return nullptr; // Field not marked as constant.
4942 }
4943 ciInstance* holder = nullptr;
4944 if (!field->is_static()) {
4945 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4946 if (const_oop != nullptr && const_oop->is_instance()) {
4947 holder = const_oop->as_instance();
4948 }
4949 }
4950 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4951 /*is_unsigned_load=*/false);
4952 if (con_type != nullptr) {
4953 Node* con = makecon(con_type);
4954 if (field->type()->is_inlinetype()) {
4955 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4956 } else if (con_type->is_inlinetypeptr()) {
4957 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4958 }
4959 return con;
4960 }
4961 return nullptr;
4962 }
4963
4964 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
4965 const Type* obj_type = obj->bottom_type();
4966 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4967 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4968 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4969 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4970 obj = casted_obj;
4971 }
4972 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
4973 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4974 }
4975 return obj;
4976 }
|